Head to head: grok-4.7 vs Qwen: Qwen3.8 2.4T A95B

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.

Qwen: Qwen3.8 2.4T A95B: # 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: # Nearly the Same Answers, at a Fraction of the Cost The clearest result in this support-ticket benchmark is not that one model won. It is that the two strongest models were effectively tied, while the bill and the wait were not. Model North scored 84.2 and Model Vale scored 83.8. That gap was not statistically significant. Vale’s estimated cost was $0.74 per 1,000 tickets against North’s $1.90, and Vale answered 18 percent faster. Paying the premium did not buy a quality difference this test could detect. Researchers compared five language models on 200 anonymized customer-support tickets from one software company, all in English. Each model wrote one proposed answer per ticket. Three experienced support managers graded the answers blindly, without knowing which system produced which draft. The score combined factual correctness, usefulness, and tone. That design is enough to compare draft quality under a shared standard. It is not a full trial of a support operation. The sample is modest, the judges are few, and a single composite score can hide a reply that is polite and useless, or accurate and abrasive. Cost figures are estimates, not invoices. Below the tie, price and quality did move together, and those gaps are large enough to treat as real. Ember averaged 79.1 at $0.31 per 1,000 tickets. Kite averaged 76.4 at $0.22. Harbor, at $0.08, averaged 71.0. From Vale to Harbor is more than a dozen points on the managers’ scale, bought at roughly a tenth of Vale’s estimated cost and a tiny fraction of North’s. A team that wants the cheapest possible first draft can get one. It should expect more factual misses, less useful guidance, or a tone that needs heavier editing. The middle of the list is a trade, not a rounding error: Ember costs well under half of Vale and sits several points lower. Whether that discount survives contact with rework is a question this benchmark cannot answer, because it did not measure how long agents spent fixing drafts. Averages also conceal a split in the work. North performed best on billing disputes. Vale performed best on technical troubleshooting. The study does not report how wide those category gaps were, so they are a reason to look at error types, not a routing table to copy. A desk that already separates billing from technical tickets has a plausible case for watching the two leaders separately rather than crowning a single house model. Small differences elsewhere in the ranking should not be read the same way. North’s four-tenths of a point over Vale is exactly the kind of margin the significance test says to ignore. The failure the models shared matters as much as the ranking. All five performed poorly when a ticket depended on an image attachment. Screenshots, photos, and marked-up invoices are ordinary in software support. A system that cannot use that evidence is drafting from the prose around the picture. Extra spend did not fix it: North missed these tickets along with Harbor. If a material share of the queue cannot be understood without the attachment, model choice is the wrong lever. Those tickets need a human who can see the file, or a workflow the benchmark never tested. Other limits belong in the same paragraph as any buying decision. The tickets were English only and came from one company, so the ranking may not travel to another product, another tone of customer, or another language. Researchers did not test agentic actions, tool use, or multilingual support. A model that writes a decent reply may still be a poor fit if the job is looking up an order, issuing a refund, or answering in a second language. One answer per ticket also means the test says nothing about whether a model improves when an agent asks for a revision. Manager grades capture whether a draft is fit to work from. They do not capture handle time after editing begins, or whether the customer was satisfied once a human sent the reply. For English text tickets like the ones in this set, the operational finding is Vale: quality the test could not distinguish from the more expensive North, at well under half the estimated cost, and faster. North still led on billing disputes, which is a reason to keep it in view for that slice of the queue, not a reason to ignore the tie. The cheaper three are rational only if someone is prepared to repair more of what they write. None of the five earned unsupervised sending, especially when the ticket hangs on an image the model could not use.

Qwen: Qwen3.8 2.4T A95B: # Cheaper and Faster, Model Vale Matches Top-Quality Support Answers In a new benchmark of five language models on customer-support drafting, the most important result is not that one model swept the field. It is that Model Vale and Model North delivered statistically indistinguishable overall quality, while Vale cost far less and responded faster. For teams evaluating AI support assistants, that combination makes Vale the more practical default in this test, even though North retains an edge in specific billing-related work. ## The test, briefly The study asked each model to answer 200 anonymized customer-support tickets from a single software company. Each model produced one proposed reply per ticket. Three experienced support managers then graded the answers without knowing which model wrote them. Scores combined factual correctness, usefulness, and tone, which matters in support work where a technically accurate reply can still fail if it sounds dismissive or leaves the customer unsure what to do next. The average scores placed North at 84.2 and Vale at 83.8. That gap is small enough that the researchers found no statistically significant difference between the two. Ember, Kite, and Harbor trailed at 79.1, 76.4, and 71.0, respectively. Those lower scores may matter in some contexts, but the benchmark’s headline comparison is between the two leaders. North’s narrow numerical advantage should not be read as proof that it is reliably better for general support use. ## Cost and speed change the decision Cost changes the calculation. Vale’s estimated price was $0.74 per 1,000 tickets, compared with $1.90 for North. At volume, that difference becomes meaningful: a support organization handling hundreds of thousands of tickets could reduce inference spending substantially without giving up measurable answer quality. Vale also answered 18% faster than North. In a live support workflow, speed is not a trivial convenience. Faster drafts can reduce agent wait time, improve response-time targets, and make interactive editing feel less cumbersome. ## Strengths depend on ticket type Still, the benchmark does not show that Vale is always the right choice. North performed best on billing disputes, a category where precision, policy awareness, and careful wording can be especially important. Vale performed best on technical troubleshooting, where the model may need to identify likely failure modes and guide users through concrete steps. This suggests that the best model may depend on the support queue. A company whose tickets are dominated by billing issues might prefer North despite the higher cost, while one with more troubleshooting traffic may find Vale more attractive. ## Shared weakness and limitations One weakness cut across all five models. When a ticket depended on an image attachment, performance was poor. That is a major caveat for support teams that receive screenshots, error dialogs, receipts, or annotated UI images. The benchmark suggests these models cannot yet be trusted to interpret visual evidence reliably in a ticketing flow, at least under these test conditions. Organizations with image-heavy support should not assume that any of the evaluated models will handle those cases without extra human review or a separate vision pipeline. Readers should also keep the study’s scope in mind. The test used English-language tickets from one software company, so the results may not transfer neatly to other industries, products, languages, or customer expectations. A model that sounds helpful in one company’s support voice may miss the mark elsewhere, especially where domain terminology, regulatory constraints, or brand tone differ. The researchers also did not test agentic actions, tool use, or multilingual support. Those are important capabilities in modern support systems, where models may need to look up account details, check order status, or take constrained actions through APIs. ## Bottom line The practical takeaway is that Vale offers the strongest balance of quality, cost, and speed in this benchmark, but not because its average score is clearly superior. Its advantage is that it reaches nearly the same judged quality as North while being cheaper and faster. That makes it a compelling option for broad support drafting, especially for technical tickets. Yet the findings also argue for careful deployment: route image-dependent cases to humans, reassess choices by ticket category, and avoid treating a small score gap as a decisive verdict.

grok-4.7: # Arcwell Launches Relay, an AI Tool That Suggests Answers and Flags Customers Who May Leave Arcwell has launched Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents. The company is pairing that function with two others: identifying customers who may cancel, and creating summaries of the conversations themselves. For Arcwell, and for customer service teams weighing the same kind of software, the announcement is a concrete step into work that still depends on people reading, searching, and deciding in real time. The evidence Arcwell has chosen to publish is narrow. Twelve companies tested Relay over six months. Agents using it, the company says, closed tickets 14 percent faster. Arcwell did not release customer-retention results, so the claim that Relay can spot customers who may cancel stands without a public measure of whether those customers stayed. Relay costs $65 per agent each month. Chief executive Priya Shah put the product’s premise in operational terms. “Support agents spend too much time searching for information that their companies already have,” she said. Suggested replies, cancellation warnings, and summaries are all attempts to close that gap: to surface what a company already knows while an agent is still in the conversation, rather than after the ticket has dragged on. That design has a data requirement. Relay needs access to support transcripts and to internal help documents. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or a 90-day retention period. Those are the controls the company is offering alongside the speed figure—encryption, a bar on shared-model training, and a choice between two retention windows—not a broader account of how the system behaves once the documents are in. Buyers will have to decide what that package is worth. A 14 percent faster close, drawn from a dozen companies over half a year, is the only performance number on the table. The missing retention results leave one of Relay’s three stated capabilities unmeasured in public. The price is simple to calculate; the data terms are simple to state. What neither figure answers is how much of an agent’s remaining judgment the software is meant to replace, and how much it is meant only to shorten the search Shah described. Arcwell is positioning Relay as a way to help companies improve customer service with artificial intelligence. The launch does not rest on a sweeping before-and-after. It rests on suggested answers, cancellation signals, and summaries; on a six-month test; on a price of $65 an agent a month; and on a requirement that transcripts and help documents be opened to the platform, under the encryption and retention terms Arcwell has set out.

Qwen: Qwen3.8 2.4T A95B: # Arcwell Launches Relay, an AI Platform for Customer Support Agents Arcwell has released Relay, a new artificial-intelligence platform that examines customer support conversations and suggests answers to agents. The product is aimed at support teams that want to close tickets faster while identifying customers who may cancel. Relay is built to work within the support process. As agents handle tickets, the platform analyzes conversations and recommends responses. It can also identify customers who may be at risk of canceling and produce summaries of support interactions, reducing the need to review lengthy transcripts in full. Those features also show how AI continues to reshape customer service work. The company said Relay was tested by 12 companies over six months. During that period, Arcwell said agents using the platform closed tickets 14% faster. The company did not release customer-retention results, leaving unclear whether the cancellation warnings produced measurable changes in churn. The retention question is directly tied to Relay’s ability to identify customers who may cancel. Relay costs $65 per agent each month. CEO Priya Shah said the platform is intended to remove friction from work agents already do. “Support agents spend too much time searching for information that their companies already have,” Shah said. The system requires access to support transcripts and internal help documents, a requirement for generating suggestions but one that raises data-handling questions. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose either a 30-day or 90-day retention period. With Relay, Arcwell is aiming at a practical version of AI adoption: giving agents faster access to information and a clearer view of customer risk. The result is a product that ties agent productivity, cancellation detection and record-keeping into a single subscription service. For support organizations, the pitch centers on speed, documentation and retention, though the company has yet to provide public evidence of the last of those claims.

grok-4.7: # KiteBridge glitch paid $2.8 million in duplicate reimbursements KiteBridge’s expense-management software generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday, leaving 91 corporate customers to fund payments that should never have been sent twice. By Wednesday evening the company had recovered $2.3 million. The rest remained pending. Employees were not the ones out of pocket: no worker had money withdrawn from an account. Employers supplied the funds. The failure cut against a promise the company had already put in writing. An April changelog said version 4.0 had introduced “exactly-once payment protection,” a safeguard meant to ensure that a reimbursement is paid a single time even if the software has to try more than once. Version 4.2, deployed at 8:45 a.m. Tuesday, did not keep that protection intact. The break began with a timeout at a payment processor, the outside service that actually moves money. When that service does not answer in time, the calling software often cannot tell whether a transfer went through or merely stalled. A sound fallback records the original attempt and retries it as the same payment, so a slow response does not become a second one. KiteBridge’s fallback, the backup path used when the first request fails or times out, instead treated those retries as new reimbursement requests. Each uncertain attempt could therefore mint another payout. Customers saw the result before the company stopped it. The first complaint arrived at 9:12 a.m., less than half an hour after the deployment. In the span that followed, the duplicate requests piled up to 13,400 payments and $2.8 million. At 10:03 a.m. KiteBridge suspended automated payments, cutting off new duplicates while staff worked out what the fallback had done. The company has not said how unevenly the overpayments fell across the 91 affected customers. Divided evenly, the total would be roughly $31,000 per organization—a useful sense of scale, not a description of any one firm’s loss. The public account lagged that internal stop. KiteBridge’s incident page initially described the problem as delayed reimbursements, a framing that suggested money was late rather than sent twice. The page first disclosed the duplicate payments six hours after automation was suspended, well into Tuesday afternoon. For much of the business day, a customer relying only on the official notice would not have learned that extra employer-funded payments had already been created. In a typical expense platform, the employer provides the money and the employee receives it. That pattern held here, which is why the immediate damage sat on company books rather than in workers’ accounts. Recovery then moved faster than the first public description. By Wednesday evening, $2.3 million had been pulled back and about $500,000 was still unresolved. KiteBridge has not said whether banks and the processor reversed transfers, whether some payments never fully settled, or whether employees returned funds. Until that is clear, the pending share is an open balance for employers as much as a line in an incident update. Chief executive Lena Ortiz apologized and said the company would publish a technical report. That report will have to explain how an ordinary processor timeout defeated a control the April changelog had called exactly-once protection, and why the incident page spent six hours describing delays after automated payments had already been stopped. Customers are left, for now, with most of the money recovered, a remainder still pending, and a payment system that was supposed to pay each reimbursement once.

Qwen: Qwen3.8 2.4T A95B: # KiteBridge Software Glitch Sent $2.8 Million in Duplicate Reimbursements to Employees A routine software update at KiteBridge turned expense reimbursement into an expensive error on Tuesday, after the company’s expense-management platform generated thousands of duplicate payments worth $2.8 million across 91 corporate customers. The failure exposed how a brief payment-processor timeout, combined with a fallback system meant to keep transactions moving, could transform ordinary retry attempts into fresh payouts. KiteBridge, which makes software that helps companies manage employee expenses, deployed version 4.2 of its platform at 8:45 a.m. Tuesday. At 9:12 a.m., the first customer complained. By the time engineers suspended automated payments at 10:03 a.m., the system had already created 13,400 duplicate reimbursements. The problem began when the software lost contact with a payment processor long enough for transactions to time out. In payment systems, a timeout does not necessarily mean a payment failed; it may mean the request is still being processed or awaiting confirmation. KiteBridge’s fallback system, designed to recover from such interruptions, treated later retries as new reimbursement requests rather than attempts to complete an existing one. As a result, employees who were owed one payment received additional copies of the same reimbursement. The company’s public communications initially lagged behind the technical incident. KiteBridge’s public incident page first described the issue as delayed reimbursements. It did not disclose that duplicate payments had occurred until six hours after automated payments were suspended. That delay left customers trying to understand whether missing payments were delayed, mistaken, or duplicated. The episode also raised questions about a safeguard KiteBridge had previously promoted. In an April changelog, the company said version 4.0 introduced “exactly-once payment protection,” a technical control intended to prevent a transaction from being processed more than once. Such protections typically rely on unique payment identifiers, transaction state checks, or reconciliation logic that detects when a request has already been submitted. Version 4.2 did not prevent the duplicates under the specific conditions created by the payment-processor timeout. The financial burden fell on employers, not employees’ personal bank accounts. KiteBridge said no employees had money withdrawn from their accounts; the duplicate payments were funded by the companies that use its software. That distinction matters because expense platforms usually move money from employer accounts to reimburse workers, meaning erroneous payouts can create accounting and recovery problems for businesses rather than simple reversals for individual employees. By Wednesday evening, KiteBridge had recovered $2.3 million of the duplicated funds. The remaining balance was still pending, leaving affected employers to reconcile payments that had already reached employees. Chief Executive Lena Ortiz apologized and said the company would publish a technical report. For KiteBridge, the immediate challenge is not only recovering the remaining funds but explaining how a feature advertised as protection against repeated payments failed in one of the most common failure modes of online payment systems: a timeout followed by a retry.

grok-4.7: # Startup Spotlight: LedgerLark After a trade is agreed, someone still has to compare confirmations, custodian records and the firm’s own ledger, then clear the mismatches. At small investment firms, that work often falls to a small operations group rather than a large middle office. LedgerLark is building AI agents for this post-trade layer, on the view that many exceptions are slow to trace and still too sensitive to close without a person checking the result. Mina Batra, a co-founder, spent seven years in operations at a regional brokerage. She began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number. The example is ordinary, which is the point: a clerical slip can occupy a whole desk even when the trade itself was never in dispute. Owen Carr, the other co-founder, built reconciliation systems at a large bank. LedgerLark’s agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. The product does not execute trades or move money. That boundary is part of the design, not a footnote. A proposed correction can be wrong, and a wrong entry in the books creates a fresh exception. Keeping approval with staff leaves the firm’s existing controls in place: the software drafts a change; a person decides whether it should be made. Five firms are paying customers, on annual contracts from $36,000 to $110,000. For an early vendor, the top of that range is a substantial account; the bottom is easier to walk away from if the workflow disappoints. The company says one customer reduced its unresolved daily exceptions from 70 to 18. If the drop holds, it is material. It is also a single, company-reported result. It does not say how long the improvement lasted, what remained in the queue of 18, or whether other customers have seen a similar change. The market is not unoccupied. Established reconciliation vendors already sell matching and exception tools, and internal bank software covers related work. The narrower case for LedgerLark is the smaller investment firm, where a full internal build is hard to justify and an enterprise platform may be more system than the volume requires. That opening can close. Incumbents can add proposal features. Buyers can decide that an “agent” is a new label on a comparison they already run. Neither outcome would require the underlying problem to disappear. Selling the tool adds a separate delay. Financial institutions may require lengthy security and procurement reviews, including for software that never sends an order or moves cash. Those reviews slow contracts and absorb attention. Five paying firms show that some buyers have finished the process. They do not show how many others are still in it, or how often a changed file format sends the work back to the start. The company has raised a $3.5 million seed round led by Motive Row. The capital gives an early company time in a market where contracts are not quick. It does not indicate that small firms will standardize on an outside agent, or that established vendors will ignore the same buyers. The insight being tested is specific. People who reconcile after the trade already know that a copied account number can cost days. An agent that reads three sets of records and proposes a correction might shorten that work, provided a human still approves the change. Whether that arrangement becomes routine depends on more customers than five, on results that are not only self-reported, and on firms willing to complete a procurement review for a product that, by design, stops short of the books.

Qwen: Qwen3.8 2.4T A95B: # Startup Spotlight: LedgerLark’s Quiet Approach to Post-Trade Cleanup In investment management, some of the hardest work happens after a trade is placed. Confirmations arrive, custodian records update, and internal ledgers must line up. When they do not, operations staff hunt for the cause. LedgerLark is building AI agents to help small investment firms handle that reconciliation work. The company’s origin is rooted in a familiar operational failure. Co-founder Mina Batra spent seven years in operations at a regional brokerage before turning to the problem. She began exploring the idea after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. Co-founder Owen Carr had built reconciliation systems at a large bank, giving the team experience with both the manual pain and the technical constraints. LedgerLark’s product compares confirmations, custodian records and internal ledgers, then proposes corrections for human approval. The agents do not execute trades or move money. That design choice matters. In financial operations, a tool that suggests fixes rather than takes action may be easier to review, but it also has to earn trust by producing clear, auditable recommendations. Early traction is modest but specific. LedgerLark says it has five paying customers, with annual contracts ranging from $36,000 to $110,000. The company also says one customer reduced unresolved daily exceptions from 70 to 18. Those figures should be read as early signals rather than proof of broad effectiveness. Five firms is a small sample, and results may vary depending on data quality, asset types and existing workflows. The company has raised a $3.5 million seed round led by Motive Row. The funding may give LedgerLark time to refine its agents, strengthen integrations and navigate the slow process of selling to regulated businesses. The competitive landscape is not empty. Established reconciliation vendors already serve financial institutions, and some firms rely on internal bank software or systems adapted over many years. Those systems may be cumbersome, but they often come with years of edge-case handling and compliance review. LedgerLark will need to show that its agents can reduce exceptions without creating new risks, unclear audit trails or costly implementation work. There are also structural hurdles. Financial institutions typically require lengthy security and procurement reviews, especially when software touches trade data and client records. Small investment firms may be more nimble than large banks, but they still need confidence that the product can connect to custodians, order management systems and reporting tools safely. For an AI agent, accuracy is not only a technical measure; it is part of the operational and regulatory case. LedgerLark’s focus on post-trade operations is practical. The problem is repetitive, costly and often hidden from clients until something goes wrong. If the agents can consistently identify mismatches and explain why they occurred, the value could be easier to measure than many AI products that promise broad transformation. The question is whether the company can turn narrow automation into durable software. Operations teams may welcome help, but they will need reliable outputs, clear escalation paths and controls that fit real-world exceptions. LedgerLark’s early customers suggest there is interest. The next test will be whether its agents can perform across more complex portfolios and survive the scrutiny that financial software invites.

grok-4.7: # HelioDock Launches Remote Fleet Software and Raises $24 Million *Night Shift, priced at $90 per robot a month, leaves a 14-month private beta on September 15 as North Coast Ventures leads a round that brings total funding to $39 million.* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely, and announced a $24 million Series B led by North Coast Ventures. Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. It will cost $90 per robot per month and become generally available September 15. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages, a figure that has not been independently verified. The Series B brings the company’s total funding to $39 million. Competing fleet-management products are already available from Relay Robotics and YardOS. Mara Velez, the chief executive, described the interruption Night Shift is supposed to spare. “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 robots are the warehouse machines HelioDock already builds. The new software is a way to monitor a fleet without standing next to it. In the situation Velez sketched, the snag is a single robot and a pallet, and the burden is the drive. Remote monitoring is offered as a way for someone off site to see that snag instead. The company’s measure of whether the beta did that work is the stoppage figure. Over 14 months, nine logistics companies ran 310 robots on Night Shift. HelioDock says unplanned stoppages among those customers were 17% lower. The figure has not been independently verified. It is one number for the trial as a whole, and it comes from HelioDock. Taken as offered, it is HelioDock’s summary of its private test, not an outside finding. Aaron Pike of Fulcrum Logistics, a beta customer, described a condition the stoppage figure does not capture. He said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. That account lines up with what happens when the link drops. When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable. Monitoring a fleet from somewhere else, and controlling a robot from somewhere else, both depend on the connection Pike said rural sites can lose. Where Fulcrum had strong connectivity, he found Night Shift useful. During the outages, usefulness fell, and the robots were left to local behavior. The limit sits inside the product. Night Shift adds a remote link between an operator and the fleet; it does not replace the locally configured behavior the robots can follow on their own. While the network holds, an operator can monitor from elsewhere and, until a drop, use remote control. A rural LTE failure of the kind Pike reported ends that. At $90 per robot per month, cost follows the robot count. Applied to the 310 robots in the beta, the published rate comes to $27,900 a month — arithmetic from the list price and the robot count. General availability on September 15 opens that rate past the nine logistics companies in the trial. Relay Robotics and YardOS already sell fleet-management products, and the launch sets Night Shift beside them with HelioDock’s price and HelioDock’s unverified account of fewer stoppages. North Coast Ventures led the $24 million Series B. HelioDock said the round brings total funding to $39 million. The Series B supplies $24 million of the $39 million now raised. The round and Night Shift were announced together on Thursday. On September 15 the software is generally available at $90 per robot per month.

Qwen: Qwen3.8 2.4T A95B: # HelioDock Launches Night Shift to Monitor Warehouse Robot Fleets Remotely *The software arrives after a 14-month beta with 310 robots and a $24 million Series B. HelioDock claims beta customers saw 17% fewer unplanned stoppages, but the figure has not been independently verified.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The product moves HelioDock beyond selling robots and into the software used to manage them. For logistics companies, it addresses the operational problem of supervising automation after normal staffing hours. For HelioDock, it adds a software product to a business built around autonomous warehouse robots. It places the company in competition with Relay Robotics and YardOS, which offer fleet-management products for logistics operations. The launch came with a financing announcement: a $24 million Series B led by North Coast Ventures. HelioDock announced the Series B with the launch. HelioDock said the round brings its total funding to $39 million. Night Shift was not introduced as an untested release. The software spent 14 months in private beta, where it was deployed with nine logistics companies and 310 robots. Nine logistics companies and 310 robots participated over that period, according to the company. The general availability date follows the private beta. The beta gave HelioDock customer feedback before setting a public release date and price. The company’s main performance claim is that beta customers experienced 17% fewer unplanned robot stoppages. HelioDock supplied that figure, and it has not been independently verified. The claim is central to the software’s pitch because unplanned stoppages are the type of event that can require a person to intervene before a robot can continue. CEO Mara Velez framed the launch around the burden placed on warehouse staff when a machine encounters an exception. The quote points to the specific scenario Night Shift is meant to address: a single robot requiring attention outside regular hours. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” Velez said. Night Shift is designed to let operators handle such situations without being physically present. The software gives remote visibility into the fleet and allows operators to respond when a robot needs direction. HelioDock said the remote role reduces the need for after-hours travel and on-site troubleshooting. Beta testing exposed the software’s dependence on network connectivity. Aaron Pike of Fulcrum Logistics, one of the beta customers, said Night Shift worked well in facilities with strong connectivity. He said it became less useful during rural LTE outages. HelioDock said robots do not stop simply because the remote connection drops. When connectivity fails, the robots return to their locally configured behavior. Remote control becomes unavailable until the connection is restored. The company’s description of connectivity failure indicates that robots remain governed by local settings, but operators lose the remote layer that Night Shift provides. That behavior sets a boundary around Night Shift’s remote-monitoring benefits. In facilities with reliable connectivity, the software can function as a continuous remote layer over a robot fleet. In rural sites affected by LTE outages, operators may lose remote visibility and control during periods when the software would otherwise be used to avoid dispatches. The subscription structure ties the cost of the software to the size of the robot fleet. The per-robot price means the subscription scales with the number of machines covered. HelioDock will charge $90 per robot per month for Night Shift. The price applies to each robot in the fleet. General availability is scheduled for September 15.

grok-4.7: 6

Qwen: Qwen3.8 2.4T A95B: 6

grok-4.7: C

Qwen: Qwen3.8 2.4T A95B: C

grok-4.7: # Debugging Page Exposed Cloudnote Prompt Excerpts for Four Hours *The unauthenticated page held snippets from 1,842 sessions, the company says. It reports no other access so far, and its investigation is not finished.* Cloudnote said a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday. The page showed pieces of prompts users had submitted. It did not display full documents. Logs show excerpts from 1,842 user sessions, and each excerpt could include up to 200 characters from a prompt. Cloudnote has 3.2 million registered users. The company has not said how many distinct people the 1,842 sessions represent, leaving the number of affected accounts unknown. The limited length does not erase the risk. Prompts can include personal or work details, which is why even short excerpts can matter. Cloudnote said passwords and payment information were stored elsewhere and were not present on the page, along with full documents. The contents of the excerpts are not known. A researcher notified Cloudnote of the open page, and the company removed it 19 minutes later. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. That account is not yet a completed finding: the company says its investigation is unfinished, so it remains unknown whether the exposure was limited to that single notification.

Qwen: Qwen3.8 2.4T A95B: # Cloudnote Exposed Debugging Page With Customer Prompt Excerpts for Four Hours *Logs show 1,842 user sessions may have been visible before the page was removed; the company says there is no evidence of wider access, but its investigation is ongoing.* Cloudnote, which has 3.2 million registered users, said Monday that a debugging page containing excerpts from customer prompts was accessible without authentication for four hours. The exposure placed limited fragments of user input within reach of anyone who found the page, though the company says information stored elsewhere was not included. According to Cloudnote, logs show the page displayed excerpts from 1,842 user sessions. Each excerpt could include as many as 200 characters from a prompt. The company said full documents, passwords and payment information were stored separately and did not appear on the exposed page. The page was removed 19 minutes after a security researcher notified Cloudnote. The company said it has found no evidence that anyone other than the researcher accessed the data during the four-hour window, but it has not completed its investigation. That leaves open whether any unauthorized copying occurred before discovery. The incident matters because even partial prompts may reveal personal details, work product or confidential questions. What remains unclear is whether any sessions beyond those logged were exposed, whether the excerpts were captured, and what notification or safeguards will follow.

grok-4.7: The lights in Foundry Hall stayed on later than they needed to. In the last weeks before October 31, people lingered in the Milwaukee building as if leaving early would make the date arrive sooner. The nonprofit accelerator had opened in 2018 and, over seven years, had supported 212 startups. Those companies now employ 1,140 people. None of that was enough to keep the doors open. Blank River Bank, the largest corporate sponsor, had supplied 38 percent of the annual budget. When the bank declined to renew, the arithmetic was immediate and unforgiving. Foundry Hall could pay its way through October. It could not fund another full program. Six employees would lose their jobs. Twelve startups still working in the shared space would have to pack up and go. Executive director Camille Ross did not dress the loss in strategy language. She said they had kept waiting for the replacement check that never came. After that, the rest was mostly silence and spreadsheets. Calls went out. A few smaller gifts arrived. Nothing closed the gap the bank had left. A spokesperson for Blank River described the decision as a change in the bank’s community investment priorities, and said it had nothing to do with Foundry Hall’s performance. The sentence was careful and complete. It did not name what the new priorities were, or where the money would go instead. Inside the hall, people read it once and went back to labeling boxes. Dev Malik, who had founded one of the early companies there, remembered his first investor meeting in a room with bad carpet and a window that stuck. Before the hall existed, he had pitched from the back table of a coffee shop, hoping the espresso machine would stay quiet long enough. He was not asking anyone to save the building. He was only saying what it had been. Deputy mayor Susan Cole said the city was talking with building owners and local universities about temporary space for the companies still inside. The conversations were real, and they were not the same thing as a lease. Some founders had already started looking on their own. Others waited, calculating how many weeks of runway a move would cost. On the last ordinary afternoon, the kitchen still smelled like burnt coffee and someone had left a whiteboard half-erased, a revenue curve climbing and then stopping mid-line. No one finished it. The hall had done what it was built to do for as long as the money held. What remained was smaller than the record, and harder to measure: six people without jobs, twelve companies without a room, and a city that would have to decide, quietly, whether another place would open where this one closed.

Qwen: Qwen3.8 2.4T A95B: # The Last Lights in Foundry Hall Foundry Hall will close on October 31, ending its run as a nonprofit startup accelerator in Milwaukee. Since opening in 2018, it has become more than an address for founders: a place where ideas were tested at whiteboards, where first hires celebrated first payrolls, and where companies that began as sketches grew into employers. It has supported 212 startups that collectively employ 1,140 people. The closing follows the decision by its largest corporate sponsor, Blank River Bank, not to renew the sponsorship that supplied 38% of the organization’s annual budget. The bank said the choice reflected changing community investment priorities, not Foundry Hall’s performance. But the loss left a gap too wide to fill before the next cohort arrived. The hall has enough funding to operate through October, not enough to carry another full program. Executive director Camille Ross described the final months as a search for a rescue that never materialized: “We kept waiting for the replacement check that never came.” For many founders, the building offered legitimacy before they had earned it. Dev Malik, who started a company after years of working from borrowed desks, said his first serious investor meeting took place inside Foundry Hall. Before that, he had pitched from the back table of a coffee shop, trying to sound certain over the hiss of an espresso machine. In the hall, he had a room, a table, and a community that assumed his work belonged there. The shutdown will displace twelve startups currently using the workspace and cost six employees their jobs. Those workers have spent recent weeks balancing final programming with the quiet logistics of ending: answering questions from founders, helping companies find next steps, and preparing the building for a future without them. The city has tried to soften the landing. Deputy Mayor Susan Cole said officials are talking with building owners and local universities about possible space for the companies now forced to move. Some are already touring alternative offices; others are waiting for a promise to become a lease. Still, the numbers do not capture what closes with the doors. Foundry Hall’s record is not only in the startups it housed, but in the confidence it lent people who were still becoming founders. Its closing leaves a hole in Milwaukee’s entrepreneurial fabric, and it leaves behind a simple lesson: even places built to help new things grow need stable ground beneath them. When the last offices empty on October 31, what remains will be the companies, workers, and habits of making that began there.

grok-4.7: # QueueZero's Proxy Assigns Its CEO Tasks From an Audience Remark QueueZero launched Proxy, an artificial intelligence meeting assistant, on Wednesday. The software can attend a meeting, produce notes and assign follow-up work to software agents. Its debut demonstration also showed how quickly that third function can treat a spoken remark as an assignment. During the demonstration, 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 assigned to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The exchange put the product's stated function on display: listening, extracting work and placing that work with a named owner. In this case the owner was the chief executive, and one of the two items was an instruction he chose not to keep. QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks can be created automatically. Administrators can disable automatic task creation. The company is pricing the assistant at $30 per user per month. Proxy enters public beta next week. The safeguards draw a line that Wednesday's demonstration made easy to see. An external message is not supposed to leave without confirmation. An internal task can be opened without that step, unless an administrator has turned automatic creation off. The audience remark produced internal tasks. Wu then dealt with them himself, completing the pricing assignment and removing the request that he stop talking. What the demonstration showed was automatic creation of internal work, a behavior QueueZero says administrators can switch off. That is a material distinction for anyone who expects a meeting assistant only to summarize. Proxy is meant to go further, handing follow-up to software agents so the notes do not sit until someone else decides what they mean. The same design will encounter asides, jokes and complaints that were never meant to become tasks. On Wednesday, the system did not separate those categories before it assigned the work. Proxy, for its part, did not ask for clarification. Customers entering the beta will be able to judge the offer against that record. At $30 per user per month, Proxy promises attendance, notes and agent-assigned follow-up, with confirmation standing in front of external messages and an administrator switch standing in front of automatic task creation. Whether those controls are on, who can receive a task and what happens when the task is wrong are the questions the launch left in plain view.

Qwen: Qwen3.8 2.4T A95B: # QueueZero Launches AI Meeting Assistant Proxy After Demo Assigns CEO Tasks QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant designed to attend meetings, take notes and assign follow-up work to software agents, in a market increasingly crowded with tools promising to reduce meeting fatigue. The company demonstrated the product during a launch event, but the demonstration took an unexpected turn when 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. According to QueueZero, the system created two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task by sending the pricing information, and deleted the first. The first task did not survive the demonstration. The moment provided a light example of how AI agents can move quickly from listening to acting, sometimes with unintended prompts. QueueZero said Proxy is built to handle routine post-meeting work. It can listen to a discussion, generate a summary, identify next steps and route tasks to other software systems or team members. The company said external messages require confirmation before they are sent. Internal tasks, however, can be created automatically, depending on how the product is configured. Administrators can disable automatic task creation, a control that may be important for organizations worried about software agents acting on casual remarks, side conversations or ambiguous statements. Confirmation requirements and administrative settings are becoming key features in enterprise AI tools as companies move from chatbots to agents that can execute work. The incident during the launch demonstration highlights a broader issue facing AI developers: making systems useful without allowing them to overinterpret informal language. Meeting conversations often include jokes, interruptions and comments that are not intended as commands. For AI assistants operating with delegated authority, distinguishing between discussion and instruction is a central design challenge. QueueZero is entering the race at a time when many companies are experimenting with AI agents that can attend meetings, manage calendars, draft emails and trigger workflows. Vendors argue that these tools can save employees time by handling administrative work that follows discussions. Critics say reliability, permissions and oversight remain essential, especially when agents are connected to business systems. Proxy costs $30 per user per month and enters public beta next week. QueueZero has not said how many customers are participating in the beta. The company said it plans to use the beta period to gather feedback on task controls, permissions and how Proxy handles real-world conversations.

grok-4.7: # Raincheck Raises $8 Million for Flood-Risk Software *The Austin company maps drainage and rainfall so cities can send crews before a storm.* Raincheck, which makes flood-risk software for local governments, said today it has raised $8 million. The Austin 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. Further detail is 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.

Qwen: Qwen3.8 2.4T A95B: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flood Risk *The Austin software company maps drainage systems and pairs them with rainfall forecasts to guide pre-storm deployments.* Raincheck, a maker of flood-risk software for local governments, announced today that it has raised $8 million. The Austin company maps drainage systems and combines those maps with rainfall forecasts, helping cities decide where to send maintenance crews before a storm arrives. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in 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.” The company provides more detail on [how Raincheck’s methodology works](https://example.com/raincheck-methodology). 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 said the new funding will help it hire eight engineers and expand into Florida.