Head to head: Z.ai: GLM 5.3 vs grok-4.7

Z.ai: GLM 5.3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

grok-4.7: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Z.ai: GLM 5.3: # Botched KiteBridge Update Paid Out $2.8 Million in Duplicate Reimbursements KiteBridge, a maker of expense-management software, issued $2.8 million in duplicate employee reimbursements this week after a faulty update — 13,400 extra payments across 91 corporate customers — before the company managed to halt its automated payout system. As of Wednesday evening, $2.3 million had been recovered, and the remaining roughly $500,000 was still pending. ## How a safeguard gave way The update, version 4.2, went live at 8:45 a.m. Tuesday. KiteBridge's software handles the routine corporate work of approving employees' expense claims and paying them back, and it relies on an outside payment processor — the service that actually moves the money — to complete each reimbursement. When the processor failed to respond within an expected window, a problem engineers call a "timeout," KiteBridge's software did what many payment systems are built to do: it tried again. But the company's fallback system — the backup logic meant to keep operations running when the primary route fails — treated each retry as a brand-new reimbursement request rather than a repeat of the original one. The result: the same expenses were paid again and again. By the time the money stopped flowing, the software had generated 13,400 duplicate reimbursements totaling $2.8 million. The failure carries an uncomfortable irony. KiteBridge's own changelog for version 4.0, released in April, said that update had introduced "exactly-once payment protection" — industry shorthand for a guarantee that every payment is executed a single time, neither duplicated nor dropped. Whatever that mechanism was designed to do, it did not stop Tuesday's cascade, and the company has not yet explained why. ## A fast-moving morning The first customer complaint arrived at 9:12 a.m., less than half an hour after the update went live. KiteBridge suspended its automated payments at 10:03 a.m., roughly 80 minutes after the rollout began. The company has not publicly detailed how the flaw escaped pre-release testing, or how quickly engineers traced the duplicates to the timeout-retry behavior. ## Employers, not employees, bore the cost Because expense reimbursements are funded by employers, the duplicate payments came out of corporate accounts rather than employees' pockets. No workers had money withdrawn from their accounts; if anything, some found unexpected extra deposits that they may now be asked to return. ## A softer first account KiteBridge's public incident page initially described the problem as delayed reimbursements — a milder framing than what was actually happening. It was not until six hours after automation was suspended, late Tuesday afternoon, that the page disclosed the duplicate payments. The distinction mattered to customers watching that page. A delay suggests money is slow to arrive; duplicates mean money has already left twice, and someone now has to ask for it back. CEO Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report. That report is likely to face pointed questions — about how an "exactly-once" guarantee failed in practice, and about why the company's public account took six hours to catch up with what its customers' payment ledgers already showed.

grok-4.7: # Employers absorb $2.8 million after KiteBridge pays reimbursements twice A faulty software update at KiteBridge, which makes expense-management software, generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday, leaving 91 corporate customers to fund payments their employees had already received once. No employees had money withdrawn from their accounts. By Wednesday evening, $2.3 million had been recovered. The rest remained pending. The trouble followed the deployment of version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m. What looked, at first, like a routine processing delay was a retry failure with a much larger bill. The duplicates had a specific cause. A timeout at the payment processor caused KiteBridge’s fallback system to treat retries as new reimbursement requests. In a normal payment flow, software sends an instruction to a processor, waits for confirmation, and tries again if the reply never arrives. The second attempt is supposed to finish the same payment, not start another. Here, the fallback lost that distinction. Each unanswered request was recorded as a fresh reimbursement, and the platform paid it again. Automated payments continued until 10:03 a.m., when KiteBridge suspended them. Between the morning release and that shutdown, the software issued the 13,400 duplicates. Suspending automation stopped new erroneous transfers. It did not, by itself, tell customers what had already gone out. Public notice lagged the technical response. KiteBridge’s incident page initially described the problem as delayed reimbursements. The page first disclosed the duplicate payments six hours after automation was suspended. For much of the day, customers reading the company’s own status update would have understood the failure as slowness rather than as money sent twice. The mismatch is sharper beside an earlier claim from the company. An April changelog said version 4.0 had introduced “exactly-once payment protection.” For a general reader, that phrase means an approved expense should be paid once, even if a network call fails and the software tries again. Tuesday’s failure cut against that claim: a processor timeout led the fallback to treat retries as new reimbursement requests, and version 4.2 issued thousands of second payments before the company intervened. The money moved in one direction that matters for who bears the loss. Expense platforms sit between an employer and its staff: a worker submits a cost, the employer approves it, and the software instructs a processor to send the reimbursement from company funds. Employees were not the source of the duplicate transfers, and none had money taken from personal accounts. Employers funded the extra payments, which is why recovery is a problem for corporate ledgers rather than a charge against individual workers. Of the $2.8 million, $2.3 million had been recovered by Wednesday evening. The remaining $500,000 was still pending. CEO Lena Ortiz apologized and said the company would publish a technical report. That report is the fuller account customers do not yet have: how a processor timeout and a fallback that treated retries as new requests produced 13,400 duplicate payments after a release deployed at 8:45 a.m., and why the incident page first called the problem a delay. Until it appears, the public record is the one the company eventually corrected — not merely delayed reimbursements, but $2.8 million paid twice, most of it back by Wednesday evening, and the rest still open.

Z.ai: GLM 5.3: # Arcwell Launches Relay, an AI Assistant for Support Agents Arcwell has launched Relay, a platform that uses artificial intelligence to assist customer support agents by analyzing conversations in real time and suggesting answers. The platform goes beyond simple reply suggestions. Relay also flags customers who appear likely to cancel, and it generates summaries of support interactions. Together, the features aim to cut down on the busywork that dominates support agents' days. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said in the announcement. The pricing is straightforward: $65 per agent per month. ## Tested Over Six Months Relay was tested by 12 companies over a six-month period. Arcwell says agents using the platform closed tickets 14% faster. The company did not release results on customer retention, leaving open the question of whether the churn-detection feature translates into measurable improvements. ## Privacy Trade-offs Relay's capabilities come with data requirements. The platform needs access to support transcripts and internal help documents to function, which means companies must weigh efficiency gains against sharing sensitive customer information with a third-party tool. Arcwell says customer data is encrypted and is not used to train shared models. Companies can also choose between a 30-day or 90-day retention period, giving them some control over how long their data remains in the system. ## A Broader Push The launch positions Arcwell in the growing market of AI tools for customer service, where companies are racing to automate and augment the work of support teams. For Arcwell, Relay marks a step toward that broader goal of helping companies improve customer service with artificial intelligence. Whether the 14% speed gain holds up beyond the pilot program — and whether companies are comfortable granting the access Relay requires — will likely determine how widely the platform is adopted.

grok-4.7: # Arcwell Launches Relay, an AI Tool for Support Agents Arcwell has launched Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents. The software can also identify customers who may cancel and create summaries of those exchanges. Arcwell is offering it as a way to help companies improve customer service, at $65 per agent each month. The case for the product, as chief executive Priya Shah frames it, is a familiar waste of time. “Support agents spend too much time searching for information that their companies already have,” she said. Relay is meant to do that searching. It reads support conversations and draws on internal help documents, then proposes a reply an agent can use rather than leaving the agent to hunt for the same material alone. The other two functions rest on the same access. Identifying a customer who may cancel depends on what the conversation shows. A summary depends on having the thread to condense, so the next person on a ticket does not have to reconstruct it. None of that works without the underlying material. Relay requires access to support transcripts and internal help documents before it can suggest an answer, flag a cancellation risk, or write a summary. Arcwell says customer data handled by the platform is encrypted and is not used to train shared models. Companies can choose how long that data is kept, selecting either a 30-day or a 90-day retention period. Those are the terms attached to the access Relay requires. The performance claim is narrower than the feature list. Twelve companies tested Relay over six months. Arcwell says agents using it closed tickets 14 percent faster. The company did not release customer-retention results from that test. A buyer can therefore see a figure for speed and nothing public on whether the customers Relay is designed to identify as cancellation risks actually stayed. That gap sits at the center of what Arcwell is selling. Suggested answers, cancellation flags, and summaries are three distinct jobs, and only one of them has a number attached. The six-month trial with 12 companies produced the 14 percent gain in how fast tickets were closed. It did not produce a published result on retention, even though spotting customers who may cancel is one of the reasons Arcwell gives for adopting the software. The price makes the tradeoff concrete. At $65 per agent each month, a support team is paying for all three functions and granting Relay the transcripts and help documents those functions depend on. In exchange, Arcwell is offering encryption, a commitment that customer data will not be used to train shared models, and a choice between 30 and 90 days of retention. What it is not offering, at least not yet, is evidence that faster tickets meant fewer customers lost. For companies weighing the launch, the record is specific. Relay examines support conversations, suggests answers, summarizes them, and tries to identify who may cancel. Agents in a half-year test at 12 companies closed tickets 14 percent faster, by Arcwell’s account. Whether that speed is worth the monthly fee, and whether the cancellation feature changes who stays, are questions the company has left for buyers to answer with less information than the product’s ambitions imply.

Z.ai: GLM 5.3: # Cloudnote Exposed Customer Prompts in Four-Hour Security Lapse *A debugging page containing excerpts from 1,842 user sessions was accessible without authentication on Monday; the company says it found no evidence anyone besides the researcher who reported it viewed the data.* Cloudnote, the note-taking and AI assistant platform with 3.2 million registered users, disclosed this week that a debugging page containing customer prompt excerpts was publicly accessible for roughly four hours on Monday. The page did not require authentication, meaning anyone with its address could have viewed the contents. Logs reviewed by the company show the page displayed excerpts from 1,842 user sessions, according to Cloudnote. Each excerpt could include up to 200 characters of a customer prompt — a fragment that may have revealed sensitive personal details depending on what users typed. The exposure was ended by an outside researcher, not Cloudnote's own monitoring: the company removed the page 19 minutes after the researcher notified it. In its disclosure, Cloudnote said it has found no evidence that anyone other than the researcher accessed the page, though it acknowledged the investigation is not yet complete. The company emphasized what was *not* exposed. Full documents, passwords and payment information were stored in separate systems and did not appear on the page, Cloudnote said. What remains unclear is how many of the 1,842 sessions involved sensitive content, whether Cloudnote can definitively rule out other visitors, and what process failures allowed an unauthenticated debugging page to reach production.

grok-4.7: # Cloudnote debug page exposed prompt excerpts from 1,842 sessions *A researcher found the unauthenticated page after it had been reachable for four hours; the company says it has seen no other access, but its review is not finished.* Cloudnote discovered that a debugging page containing customer prompt excerpts was accessible without authentication for four hours on Monday. Logs show excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters from a prompt, not the full text a customer entered. The figure is a small share of Cloudnote’s 3.2 million registered users, but the exposure still matters. Prompts can hold personal or work details, and even a short excerpt can reveal more than a customer meant to share outside the product. Full documents, passwords and payment information were stored elsewhere and were not present on the page. Cloudnote removed the page 19 minutes after a researcher notified it of the exposure. The company says it has found no evidence that anyone besides the researcher accessed the page. That statement is the company’s, and it has not completed its investigation, so outside access has not been ruled out. Until that review is done, the confirmed record is limited: an unauthenticated debugging page, a four-hour window on Monday, excerpts tied to 1,842 sessions and a cap of 200 characters. What those excerpts contained, and whether anyone else viewed them, is not yet known.

Z.ai: GLM 5.3: # QueueZero's New AI Meeting Assistant Takes Its Cue a Little Too Well *The assistant, Proxy, launched Wednesday and promptly assigned its own CEO homework.* QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to sit in on workplace meetings, generate notes, and assign follow-up work to software agents. The product enters public beta next week at $30 per user per month. But the launch demonstration delivered more than the company may have planned for. Midway through the presentation, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy took the comment literally. The assistant created two tasks and assigned both to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task — the pricing information was sent — and deleted the first. The moment, while awkward for the person giving the demo, illustrated both the promise and the pitfalls of agents that act on conversational input. Proxy's core pitch is that meetings produce commitments, and that software agents should be able to capture and execute them without a human transcribing action items afterward. QueueZero says Proxy includes guardrails around how those tasks move into the world. The assistant requires human confirmation before any external message is sent, meaning tasks that would reach customers, partners, or anyone outside an organization cannot go out on their own. Internal tasks, however, can be created automatically — a policy that made Wu the first, and presumably least enthusiastic, test case. Administrators can also disable automatic task creation entirely if they would prefer that Proxy confine itself to note-taking. The incident raises questions the industry is still working through: as meeting assistants grow from passive transcription tools into active agents, remarks made in jest, frustration, or hypothetical framing all become potential instructions. A customer joking about what a CEO should do next is harmless. Less clear is what happens when the same system hears an offhand complaint about a colleague, or a sarcastic suggestion that a project be abandoned. For now, QueueZero is treating the episode as a live demonstration of the product working exactly as designed. Whether the market agrees — and whether other executives are comfortable attending meetings where their own to-do lists can be edited by anyone with an opinion and a microphone — will become clearer when the public beta opens next week. Wu, for his part, has at least one task confirmed complete. The pricing sheet, by all accounts, was sent.

grok-4.7: # QueueZero’s Proxy Turns an Audience Remark Into Tasks for Its CEO QueueZero launched Proxy, an artificial intelligence meeting assistant, on Wednesday. The company says the software can attend a meeting, produce notes and assign follow-up work to other software agents, shifting action items out of a transcript and into a queue that agents can pick up. The launch demonstration made that handoff concrete. A customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the remark as an instruction. It created two tasks and assigned both to QueueZero chief executive Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. Only one of the two assignments left the room with him. The exchange was brief, but it showed how Proxy can treat casual speech as work to be filed rather than as commentary to be ignored. QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks can be created automatically. Administrators can disable automatic task creation, a control that would stop items like Wednesday’s pair from landing on a colleague’s list without a person approving them first. The company has drawn a line between outbound communication, which waits for confirmation, and internal assignment, which does not have to. That distinction is likely to matter more to buyers than the joke that exposed it. A tool that files work inside a company without a second look will save time when the instruction is clear and create noise when it is not. On Wednesday the assignee was in the room, heard the remark and removed the task he did not mean to keep. The same reading, applied to someone who was not present, would leave an assignment without the context that made it obvious which half was serious. Proxy is built around that reading. Notes are not the end of the meeting in QueueZero’s account of the product; follow-up is handed to software agents so a person does not have to retype the list. The demonstration showed the benefit and the risk in the same minute: a pricing request became a completed task, and an aside about talking became a task that had to be deleted. Proxy costs $30 per user per month. It enters public beta next week. QueueZero has not published a date for general availability, a list of the agents Proxy can assign or a description of how the system ranks competing remarks when several people speak. It also has not said how administrators review tasks created during a call after the fact, beyond the option to turn automatic creation off. For companies weighing the beta, the public record is a short sequence. A customer asked, in passing, for a pricing sheet and an end to the talking. The software wrote both down as work for the chief executive. He sent the sheet and erased the other item. Whether that sequence is an edge case or a standing feature of letting an assistant turn speech into assignments is the question the beta will have to answer.

Z.ai: GLM 5.3: **6** Each line violates a rule: 1. `#Flood software moves upstream` — missing space after `#` (headline format) 2. `The company is expanding its pilot.` — dek is not italicized 3. `##Details` — missing space after `##` (heading format) 4. `Read [the announcement](www.example.com/news)` — link is not an absolute HTTPS URL 5. `* The first pilot begins Monday.` — list item uses asterisk instead of hyphen 6. `<div>The second pilot begins Friday.</div>` — raw HTML is forbidden

grok-4.7: 6

Z.ai: GLM 5.3: # A Tie at the Top: What a 200-Ticket Support Benchmark Really Shows The most interesting number in this benchmark isn't a score. It's a price tag. Researchers tested five language models on 200 anonymized customer-support tickets from a single software company. Each model produced one proposed answer per ticket — no retries, no clarifying questions. Three experienced support managers graded every answer blind, never knowing which model had written what, scoring each on factual correctness, usefulness, and tone. The priciest model, North, finished first with an average of 84.2. Vale, at 83.8, was statistically indistinguishable from it. What the two did not share was economics: North cost an estimated $1.90 per 1,000 tickets, Vale $0.74. Vale also answered 18% faster. | Model | Avg. score | Est. cost / 1,000 tickets | | --- | --- | --- | | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | ## The tie is the finding A 0.4-point gap on 200 tickets isn't a ranking; it's noise you can't distinguish from zero, and the significance test says so explicitly. On this workload, premium pricing did not buy measurable quality. That's a familiar pattern in maturing technologies: once several vendors clear the quality bar, the bar stops being the differentiator. What separates the top two here is everything around the score. Vale costs roughly 39 cents on North's dollar and responds faster on top of it. The absolute dollar figures are small; the multiples are what scale with ticket volume. ## "Best" depends on your queue One caveat keeps the tie from being a simple verdict. The aggregate scores hide category-level differences: North was strongest on billing disputes, Vale on technical troubleshooting. A queue dominated by billing escalations might still justify North's premium; one dominated by troubleshooting tickets would be paying extra for a slight weakness. The tie, in other words, was measured on one company's ticket mix, in one language. Change the mix and the ranking could shuffle — the study can't say which. ## The price–quality curve bends hard Below the top pair, the economics change shape. Harbor, the cheapest model, costs about 4% of North's price and scores thirteen points lower. Ember costs less than half of Vale's rate and gives up nearly five points. Whether those trades are worth making depends on what a point of quality is worth in your operation — but the pattern is clear: most of a top model's price buys the last fraction of a point. Kite and Harbor occupy a range best described as "cheap enough to draft with, risky enough to supervise." Nothing shows a 71 is unusable, but nothing suggests you'd want those answers reaching customers unreviewed. ## Where all five failed The most operationally useful result isn't in the ranking at all. Every model performed poorly when a ticket depended on an image attachment — and screenshots are the lifeblood of most real support queues. The benchmark doesn't say how many tickets involved attachments, so the drag on these averages is unmeasured. The practical implication is blunt regardless: whatever model you buy, route image-dependent tickets to humans or to a pipeline built for the job. No vendor choice solves this yet. ## What this test can't tell you Keep the limits in view before generalizing: - **Scope.** One software company, English-only tickets. Retail, telecom, and multilingual queues are untested territory. - **One shot per ticket.** Real deployments allow retries, follow-up questions, and escalation paths that can move results. - **No tools, no agency.** Models answered from ticket text alone — no order lookups, account checks, or documentation grounding, which is where support correctness often lives. Agentic actions went untested. - **Human judgment.** Blind grading controls for bias, not subjectivity; three managers' scores are still three people's opinions. - **Estimated costs.** Real invoices depend on prompt length, volume, and vendor terms. ## The takeaway Don't default to the premium tier. The strongest signal here is that top-tier quality at this task has become competitively priced, and the real differentiators are now category fit, speed, and unsolved gaps like image handling. The right next step for any buyer is to rerun this kind of blind evaluation on their own queue, weighted by their own ticket mix. The models may be converging. Ticket queues never do.

grok-4.7: # Near-Equal Quality, Half the Cost: What 200 Support Tickets Say About Model Choice The clearest result in this benchmark is not which model scored highest. Model North edged the field with an average of 84.2, but Model Vale landed at 83.8—a gap the researchers judged not statistically significant. Vale did that work for an estimated $0.74 per 1,000 tickets, against North’s $1.90, and answered about 18 percent faster. For a support operation that cares about cost and latency as well as answer quality, the practical leader is the near-tie, not the nominal winner. ## How the scores were produced Researchers ran five language models over 200 anonymized customer-support tickets from a single software company. Each model wrote one proposed reply per ticket. Three experienced support managers graded those replies blindly, so brand names and price tags could not color the marks. The composite score mixed factual correctness, usefulness, and tone—the things a manager would actually defend to a customer. That design is strong enough to compare draft quality under a shared rubric. It is not a live trial. Models did not take actions, call tools, or handle follow-up turns. They also never saw multilingual tickets. The sample is English-only and drawn from one product environment, so the ranking describes that setting, not support in general. ## Quality does not fall in lockstep with price Below the North–Vale cluster, the drop is easier to see and more decision-relevant than the 0.4-point gap at the top. Ember averaged 79.1 at $0.31 per 1,000 tickets. Kite sat at 76.4 for $0.22. Harbor, the cheapest at $0.08, averaged 71.0. Those are not rounding errors. A team that treats every model as interchangeable will feel the difference in wrong facts, thin answers, or tone that managers would not send. The cost curve is steep at the top and flatter underneath. Moving from North to Vale cuts estimated spend by more than half without a detectable quality loss on this test. Moving from Vale to Ember saves more money again, but gives up several points that the graders could see. Harbor is an order of magnitude cheaper than North and clearly weaker. Price is a real lever; it is not a free one. Speed belongs in the same conversation. Vale’s 18 percent advantage over North is operational, not cosmetic, if queues are long and agents are waiting on a draft. The study does not report latency for the other three models, so speed should not be invented into a full ranking. ## Strengths split by ticket type Averages hide a useful split. North was strongest on billing disputes, where a wrong figure or a sloppy policy reading is expensive. Vale was strongest on technical troubleshooting, where usefulness depends on diagnosing the problem and pointing the customer at a workable next step. A company whose queue is mostly invoices and refunds has a different best fit from one whose queue is mostly setup and error messages. Blended scores still matter, but routing—or at least weighting—by ticket type is more honest than crowning a single champion. ## A shared blind spot Every model did poorly when the right answer depended on an image attachment. That is not a footnote. Screenshots of error dialogs, invoices, and settings pages are ordinary in software support. A benchmark of text replies will overstate readiness if those tickets are common in production. Until a system can reliably use the attachment, a human still has to open the image, and the model’s draft is only a partial assist. ## What a buyer should take from this On this evidence, Vale is the default draft model: indistinguishable from North on the blinded composite, substantially cheaper, and faster. North remains the better bet where billing disputes dominate and the extra cost is acceptable. Ember, Kite, and Harbor are plausible only if the organization will accept a visible quality cut—or will confine them to low-stakes tickets and keep humans in the loop. None of the five is ready to own image-dependent cases alone. The limits should travel with that recommendation. Two hundred tickets, one answer each, and three graders from one company cannot settle variance across products, languages, or seasons. Estimated cost per 1,000 tickets is not a full total-cost figure; review time, escalation, and customer harm from a bad reply can dwarf token spend. The test also stopped at proposed answers. It did not measure tool use, account lookups, or agentic actions, which is where many support deployments actually fail or succeed. Treat the result as a procurement signal, not a deployment certificate. Vale matches the top score where it counts, at a price and speed that change the business case. Confirm that pattern on your own tickets—especially billing, troubleshooting, and anything that arrives as a screenshot—before you lock the choice in.

Z.ai: GLM 5.3: # Raincheck Raises $8 Million to Help Cities Plan for Floods *The Austin startup's software combines drainage maps with rainfall forecasts so cities can prioritize fixes before storms arrive.* Raincheck, a company that makes flood-risk software for local governments, has raised $8 million in seed funding. The company, based in Austin, maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million to date. CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck's [methodology](https://example.com/raincheck-methodology) combines drainage infrastructure data with rainfall forecasting to produce its recommendations. Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations only. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

grok-4.7: # Raincheck raises $8 million for flood-risk software *The Austin company maps drainage systems and rainfall forecasts so local governments can send maintenance crews before a storm.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. Raincheck describes the approach in its [methodology](https://example.com/raincheck-methodology). The seed round was led by Basin Ventures. Existing investor Long Mile also participated. Raincheck has raised $10.5 million total. CEO Ana Flores said: > Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives. Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Z.ai: GLM 5.3: # Foundry Hall to Close After Six Years MILWAUKEE — Foundry Hall, the nonprofit accelerator that has nurtured more than 200 Milwaukee startups since 2018, will close its doors on October 31. The announcement came after Blank River Bank, the organization's largest corporate sponsor, declined to renew a sponsorship that supplied 38% of Foundry Hall's annual budget. Executive director Camille Ross said the organization held out hope longer than it probably should have. "We kept waiting for the replacement check that never came," Ross said. A spokesperson for Blank River Bank said the withdrawal reflects a shift in the bank's community investment priorities, not dissatisfaction with the accelerator's work. The distinction offers little comfort to the six employees who will lose their jobs, or to the twelve startups currently working out of Foundry Hall's shared space, all of whom must find new homes by the end of next month. The numbers behind the closing tell the story of what Milwaukee is losing. Since opening its doors six years ago, Foundry Hall has supported 212 startups that now collectively employ 1,140 people. Ross said the organization has enough funding to operate through October, but not enough to run another full program cycle — leaving leadership with the choice between a slow decline and a dignified exit. They chose the exit. For the founders who passed through its doors, the closing is more than a business story. Dev Malik, who founded a startup through the accelerator, remembers what life was like before Foundry Hall gave him a place to work and meet investors. His first investor pitch, he said, happened in that building. Before that, he was pitching from the back table of a coffee shop. That arc — from coffee shop to boardroom — is the one Foundry Hall was built to offer. It's the reason Malik and others describe the space less like an office and more like a launching pad. There may be a soft landing ahead. Deputy Mayor Susan Cole said the city is in talks with building owners and local universities about finding space for the twelve displaced companies. The effort is still in its early stages, and no commitments have been announced. For now, the founders have until Halloween. After that, the offices where 212 startups learned to pitch, hire, and grow will sit quiet — and a city that got used to having a place for its dreamers will have to decide what comes next.

grok-4.7: Foundry Hall will close on October 31. The nonprofit accelerator has worked out of Milwaukee since 2018, long enough that its absence will be felt less as a headline than as a missing room: a place where a person with an unfinished idea could sit down and be taken, for an hour, seriously. Since it opened, the hall has supported 212 startups. Those companies now employ 1,140 people, a number that will outlast the organization that helped produce it. Most of that workforce will not notice the date. The people still inside will. The closing follows one decision. Blank River Bank, the hall’s largest corporate sponsor and the source of 38 percent of its annual budget, declined to renew. A spokesperson said the choice reflected a shift in the bank’s community investment priorities and had nothing to do with Foundry Hall’s performance. The sentence was precise, and it left a gap no other pledge has filled. Executive director Camille Ross did not dress the consequence up. “We kept waiting for the replacement check that never came.” What is left will carry the organization through October. It will not carry another full program. Six employees will lose their jobs when the month ends. Twelve startups still using the workspace will have to leave it. Founder Dev Malik talked about the building in smaller terms than a budget. His first investor meeting happened there. Before that, he had been pitching from the back table of a coffee shop, hoping the room would not drown him out. He did not say the hall had made his company. He said it had given the ask a place to land. The difference will not keep the doors open, but it explains why the closing feels, to the people who used the room, like more than a lease running out. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the companies still inside. The talks are underway. They are not a lease. October is. In the weeks remaining, the hall will look much as it has: shared desks, a conference room that could be booked, whiteboards carrying half-erased plans. Someone will still take a call in the hallway. Someone will still leave a marker uncapped. The ordinary motions are what the place was for. Foundry Hall was the interval between an idea and a payroll, and that interval now has a date on it. The 1,140 jobs will not vanish with the sign. The twelve companies packing up, and the six people who kept the lights on for them, are a smaller count, and a sharper one.

Z.ai: GLM 5.3: # HelioDock Launches Night Shift Fleet Software, Raises $24 Million *The remote monitoring tool spent 14 months in private beta across nine logistics companies; pricing starts at $90 per robot per month.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, a software product that lets operators monitor and manage robot fleets remotely rather than responding to problems on site. The release arrives alongside a $24 million Series B funding round led by North Coast Ventures, bringing the company's total funding to $39 million. Night Shift will be generally available on September 15 at a price of $90 per robot per month. The launch addresses one of the more persistent pain points in warehouse automation: unplanned robot stoppages that can leave facilities shorthanded overnight. According to the company, Night Shift's 14-month private beta spanned nine logistics companies and roughly 310 robots, and participants experienced 17% fewer unplanned stoppages. That figure comes from HelioDock and has not been independently verified. The product's premise is straightforward. Robots operating in a warehouse occasionally encounter situations they cannot resolve on their own — a misaligned pallet, a blocked path, an unexpected obstacle. Traditionally, resolving those stoppages has required a human on site. Night Shift instead gives operators a remote view into the fleet, allowing them to diagnose problems and redirect robots from off site. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," CEO Mara Velez said in a statement. ## Connectivity Is the Catch The remote model has a built-in dependency: it only works while robots remain connected. Beta tester Aaron Pike of Fulcrum Logistics said the system performed well in facilities with strong connectivity but became less useful during rural LTE outages. When connectivity fails, Night Shift's remote control becomes unavailable, and robots revert to their locally configured behavior — meaning operators lose the ability to intervene remotely until the connection is restored. That limitation may shape where the product lands first. Facilities in dense areas with reliable network infrastructure stand to benefit most, while operators in more remote locations — precisely the sites where overnight staffing is hardest to arrange — may find the tool's coverage uneven. ## A Competitive Field HelioDock is not alone in the fleet-management space. Competing products are available from Relay Robotics and YardOS, both of which target the same basic problem of coordinating large numbers of autonomous machines. HelioDock's bet is that pairing fleet software directly with its own robot hardware will appeal to customers who want a single vendor across both layers. The pricing model also reflects a shift in how warehouse automation is sold. At $90 per robot per month, Night Shift is structured as an ongoing subscription rather than a one-time purchase, allowing operators to scale monitoring costs with the size of their fleet rather than committing to fixed licensing fees. ## What the Beta Showed The nine-company, 310-robot beta represents a meaningful but limited test bed. The 17% reduction in unplanned stoppages reported by HelioDock, if it holds at broader scale, would translate into real operational savings for facilities that run overnight shifts with minimal human staffing. But the figure stems from the company's own account of the beta period, and the conditions of a private trial — with hand-picked customers and close vendor involvement — may not match those of general availability. Pike's experience with connectivity gaps suggests the product's real-world performance will vary by site, and HelioDock has not detailed how Night Shift handles degraded networks beyond the fallback behavior he described. ## Pricing and Availability Night Shift becomes generally available on September 15. Pricing is set at $90 per robot per month, with the new funding round — $24 million led by North Coast Ventures — expected to support the rollout. For warehouse operators weighing the product against offerings from Relay Robotics and YardOS, the coming months will show whether the beta results hold outside controlled conditions, and how much connectivity quality determines where Night Shift delivers on its core promise: keeping robots moving without sending someone across town at two in the morning.

grok-4.7: # HelioDock Launches Night Shift Remote Fleet Software and Raises $24 Million *The warehouse robotics firm says a 14-month private beta cut unplanned stoppages, while a participating customer describes limits tied to connectivity.* HelioDock, a maker of autonomous warehouse robots, launched Night Shift on Thursday. The software is built to let operators monitor robot fleets remotely. The same day the company announced a $24 million Series B led by North Coast Ventures, an amount that brings HelioDock’s total funding to $39 million. Night Shift reaches the market after 14 months in a private beta that involved nine logistics companies and 310 robots. HelioDock says beta customers experienced 17 percent fewer unplanned robot stoppages. That figure has not been independently verified, so the reported reduction remains a company-supplied claim drawn from the closed trial rather than an outside measurement. The beta therefore supplies the only performance data the company has released, covering a defined set of logistics operators and a fixed number of machines over more than a year of private use. Chief executive Mara Velez framed the product around the practical burden of on-site intervention. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” she said. The statement isolates a specific failure mode—a robot halted by an unexpected pallet—and the resulting need for a manager to travel to the facility during overnight hours. Night Shift is positioned as the means for an operator to observe and address such events without that trip, provided the remote link stays intact. Experience inside the beta shows that the link is not guaranteed. Aaron Pike of Fulcrum Logistics, one of the nine participating companies, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His account distinguishes two operating conditions: sites where the network supported continuous remote oversight, and rural locations where LTE interruptions reduced the software’s value. The distinction matters because many warehouses sit outside dense urban coverage, and the beta included at least one customer that encountered those gaps. The software’s behavior under those gaps is explicit. When connectivity fails, robots return to their locally configured behavior and remote control becomes unavailable. Operators lose the ability to direct the machines from afar; each robot continues under the settings already stored on site until the connection returns. The fallback keeps the fleet moving according to its prior local instructions, yet it also suspends the remote monitoring function that Night Shift is meant to provide. Pike’s observation about rural LTE outages therefore lines up with a designed limitation rather than an unforeseen outage. Pricing and timing are set. Night Shift will cost $90 per robot per month and become generally available on September 15. A fleet of the size used in the beta—310 robots—would generate a monthly outlay calculated at that per-robot rate once the software leaves private testing, though HelioDock has not stated how many of the beta machines will convert to paid use. The September date marks the end of the 14-month restricted period and the start of open commercial access. Other fleet-management products already exist. Competing offerings are available from Relay Robotics and YardOS. HelioDock’s entry therefore joins an existing category rather than creating one, and operators evaluating Night Shift can weigh its remote-monitoring approach, its stated beta results, its connectivity constraints, and its $90 monthly price against those alternatives. The company has not released side-by-side data, so any comparison rests on the separate product details each vendor chooses to publish. The funding round supplies the capital context for the launch. North Coast Ventures led the $24 million Series B, and the new money lifts cumulative investment in HelioDock to $39 million. The announcement places the financing alongside the product release, linking the resources raised to a company whose robots already operate in warehouses and whose new software layer attempts to extend oversight beyond the building walls. How those funds will be allocated is not described; the disclosed facts stop at the round size, the lead investor, and the resulting total. Taken together, the Thursday announcements establish a commercial schedule, a price, a beta history, a company-reported reduction in stoppages, a chief executive’s account of the overnight problem the software targets, and a customer’s report of degraded usefulness when rural LTE drops. They also establish the boundary condition that loss of connectivity returns every robot to local behavior and removes remote control. Those elements define what Night Shift currently offers and where its remote function stops.

Z.ai: GLM 5.3: C

grok-4.7: C

Z.ai: GLM 5.3: # Startup Spotlight: LedgerLark Bets AI Agents Can Clear the Post-Trade Backlog The least glamorous corner of finance has a way of generating the most urgent problems. Trades get confirmed, custodians record them, internal ledgers log them — and then the three systems disagree, and someone's week disappears. LedgerLark, a startup building AI agents for post-trade operations at small investment firms, is betting that this particular kind of drudgery is well suited to automation. The company has raised a $3.5 million seed round led by Motive Row, and it currently counts five paying firms as customers, with annual contracts ranging from $36,000 to $110,000. ## The Insight Came From the Trenches Co-founder Mina Batra spent seven years in operations at a regional brokerage, which is to say she spent seven years watching small teams absorb the error rates of much larger counterparties. Her co-founder, Owen Carr, built reconciliation systems at a large bank — the kind of purpose-built software that big institutions can afford and smaller firms mostly cannot. The founding story, as Batra tells it, is disarmingly specific: a six-person team once spent three days resolving a single trade mismatch caused by a copied account number. Not a complex derivatives dispute or a settlement failure across time zones — a transposed set of digits, and three days of phone calls and spreadsheet archaeology to find it. That anecdote points at the gap LedgerLark is targeting. The problems aren't conceptually hard. They're numerous, fragmented, and boring, which is exactly the profile of work that human teams burn out on and that large vendors don't bother serving at small scale. ## What the Product Actually Does LedgerLark's agents compare confirmations, custodian records, and internal ledgers — the three-way match that operations teams perform daily. When the systems disagree, the agents don't just flag the exception; they propose a correction for a human to approve. That last step is a deliberate constraint. The product does not execute trades or move money. It drafts fixes and routes them to people. In a domain where a wrong correction can create a real settlement problem, keeping humans in the approval loop is both a selling point and, presumably, a hedge against the early-stage reputation AI tools have for confident mistakes. The early traction numbers are modest but suggestive. LedgerLark says one customer reduced unresolved daily exceptions from 70 to 18. If that figure holds across customers, it describes a real change in how a small operations team spends its day — though one customer's data point is one customer's data point, and the company has yet to publish anything broader. ## The Obstacles Are Honest Ones The competitive picture here is worth dwelling on, because it's genuinely crowded. Established reconciliation vendors already sell into this space, and large banks build their own internal software. LedgerLark's implicit argument is that both categories are aimed at institutions with big operations teams and bigger budgets, leaving smaller firms with manual work and spreadsheets. Whether that underserved middle really exists, and whether it will pay enough to sustain a software business, remains an open question. A five-firm customer base can't answer it. There's also a structural drag that anyone selling to finance knows well: security and procurement reviews can take months, sometimes longer than it took to build the product being reviewed. For a seed-stage company, slow sales cycles aren't an inconvenience — they're an existential design constraint. Five customers at seed-stage prices suggests LedgerLark has found a way through, or at least a wedge in. It doesn't yet suggest a repeatable motion. And the technical risk isn't zero. Reconciliation is a domain where errors are expensive and trust is slow to build. An AI agent proposing corrections is only useful if its proposals are right often enough that humans stop checking its work — and if it's wrong at exactly the wrong moment, one bad incident could undo a lot of goodwill. ## A Real Problem, an Unproven Company What's easy to appreciate about LedgerLark is the specificity. This isn't "AI for finance" in the abstract; it's agents doing a defined matching task for a defined customer profile, with a defined safety boundary. The founders' backgrounds map cleanly onto the problem — one knows the pain, the other knows the systems. What's unresolved is the standard seed-stage question: does this wedge widen? The reconciliation market's incumbents aren't asleep, the buyers are slow, and a handful of contracts proves demand more than durability. But the underlying observation — that small firms were drowning in exceptions nobody else would automate — is a real one. Batra didn't have to invent a market; she watched one from the inside for seven years. That's not a guarantee of anything. It's just a better starting point than most.

grok-4.7: # Startup Spotlight: LedgerLark A copied account number is a small error with a long cleanup. Mina Batra began exploring the idea behind LedgerLark after seeing a six-person team spend three days resolving a trade mismatch caused by one. She had spent seven years in operations at a regional brokerage. The mismatch was ordinary. The cost was the time required to find it. LedgerLark builds AI agents for post-trade operations at small investment firms. The agents compare confirmations, custodian records, and internal ledgers, then propose corrections for human approval. The product does not execute trades or move money. A proposed correction stays a suggestion until someone at the firm accepts it. Post-trade work sits close to cash and securities, and an automated step past approval would be a different product. Co-founder Owen Carr built reconciliation systems at a large bank, so the company pairs a brokerage operations background with experience on the systems side of reconciliation. The starting point is specific: a copied account number, a three-day mismatch, and the question of whether an agent can shorten that work without taking the decision away from the firm. Competitors include established reconciliation vendors and internal bank software. LedgerLark is aimed at small investment firms, where a full vendor platform may be more system than a team wants and an internal build may be out of reach. That is a plausible opening, and also a place larger vendors can follow if the accounts are worth having. A firm that already trusts its reconciler needs a reason to add another system and another review. A shorter daily exception queue might be that reason. A demonstration is not. Five firms are paying customers, on annual contracts from $36,000 to $110,000. The company says one customer reduced unresolved daily exceptions from 70 to 18. A change of that size would matter to a small team if it holds, and if the work did not simply shift into longer reviews of the agents’ proposals. The figure is the company’s, from a single customer, and it does not describe how the count was kept. Five contracts show that some firms will pay. They do not yet show how the agents behave through an audit, a volume spike, or a week when the usual reviewer is away. Financial institutions may require lengthy security and procurement reviews, and that clock is not fully LedgerLark’s to set. A system that reads confirmations, custodian records, and internal ledgers will be asked where the data sits, who can see a proposed correction, and what the firm should do when a proposal is wrong. Human approval is a real control only if the approver can check the suggestion. If that review consumes most of the time the agent was meant to save, the exception count will be a weak defense of the contract. The company has raised a $3.5 million seed round led by Motive Row. A round of that size can carry a small team through early deployments while procurement moves slowly. It does not measure the product against existing reconciliation tools, and the lead investor’s name does not either. How often is a proposed correction the right correction? How much reviewer time is left? Will established vendors or internal bank software cover the same small-firm work before the customer list grows past five? The three-day search Batra saw began with a copied account number. It is a clear picture of the waste. It is not yet a measure of whether LedgerLark’s agents would have shortened the search, or pointed it at the wrong field.