Head to head: Meta: Muse Spark Open vs DeepSeek-V4-Pro

Meta: Muse Spark Open: # Arcwell Bets on AI Agents With Launch of Relay **Support agents using the new customer-service platform closed tickets 14% faster in testing, the company says.** Arcwell has launched Relay, a new platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents in real time. The company is pitching Relay as a way to ease workload for support teams rather than replace them, cutting the time agents spend hunting for information spread across transcripts, tickets and knowledge bases. Relay is built around three core capabilities. It suggests answers to agents, flags customers who may be about to cancel, and creates summaries of support interactions. Together, Arcwell says, the features show how AI can take on some of the rote work that slows down service operations. The system was tested by 12 companies over six months. Agents using Relay closed tickets 14% faster, according to Arcwell. The company did not release customer-retention results from the test period, leaving the impact of its at-risk customer identification feature unclear. Relay costs $65 per agent each month. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. The benefits come with access requirements. To generate its suggestions and summaries, Relay requires access to support transcripts and internal help documents — the same repositories that make its answers possible in the first place. Arcwell said customer data is encrypted and is not used to train shared models, a line aimed at addressing enterprise concerns about privacy and data leakage as AI tools move deeper into customer-facing workflows. Companies using Relay can choose between a 30-day or a 90-day retention period for their data.

DeepSeek-V4-Pro: # Arcwell Debuts Relay, an AI Assistant That Reads the Room During Customer Calls Arcwell is entering the enterprise AI race with Relay, a platform that listens to customer support conversations in real time and surfaces answers, flags at-risk accounts, and drafts summaries—all without an agent having to hunt through a knowledge base. The launch marks the company’s most ambitious move into the customer service stack. Relay was shaped over a six-month pilot with 12 companies. Arcwell reports that agents using the tool closed tickets 14% faster than those relying on conventional workflows. The firm did not disclose customer-retention figures tied to the churn-detection feature. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said in a statement accompanying the announcement. Relay’s core capabilities fit into three lanes. It suggests context-aware responses drawn from a company’s own documentation, flags customers whose language or behavior signals cancellation risk, and auto-generates conversation summaries so agents spend less time on wrap-up notes. Taken together, the features sketch a vision of AI that operates as a coach and an early-warning system rather than a simple autocomplete tool. The platform arrives at a moment when enterprises are scrutinizing AI tools for both utility and privacy. Arcwell says Relay requires access to support transcripts and internal help documents to function. Customer data is encrypted and, the company emphasized, is not used to train shared machine-learning models. Clients can choose between a 30-day and a 90-day data retention window, a detail likely to draw attention from compliance teams. Relay is priced at $65 per agent per month, a rate that positions it in the middle tier of AI copilot tools. Arcwell’s ability to convert pilot interest into paid contracts will depend on whether its speed claims hold up at scale and whether companies judge the privacy architecture as rigorous enough for regulated industries. The product is available now.

Meta: Muse Spark Open: # Cloudnote Debug Page Left Prompt Excerpts Open for Four Hours *Company says a Monday lapse exposed fragments from nearly 2,000 user sessions, raising privacy concerns while investigators work to determine who saw them* Cloudnote said it discovered Monday that a debugging page displaying customer prompt excerpts had been accessible without authentication for four hours, and that it removed the page 19 minutes after a security researcher flagged it. Logs reviewed by the company show the page displayed up to 200 characters from prompts in 1,842 user sessions, a fraction of Cloudnote's 3.2 million registered users. Full documents, account passwords and payment information were not present on the page, which the company said is stored separately. The exposure still matters because prompts to the note-taking service often contain sensitive personal, financial or work details that users assume will remain private, and even short 200-character excerpts could reveal identities, health concerns, or confidential business projects if combined with other data. Cloudnote said it has found no evidence so far that anyone besides the notifying researcher accessed the page, but it has not completed its investigation and therefore cannot yet confirm who, if anyone, viewed the data or how it might have been used. Investigators are still determining whether search engines or automated scanners cached or indexed the page during the window. The company said it is reviewing access logs and outreach obligations as regulators and customers seek a fuller accounting of the incident this week.

DeepSeek-V4-Pro: # Cloudnote Left Internal Debugging Page Publicly Accessible, Exposing 1,842 User Prompt Snippets *A four-hour exposure window raises fresh questions about the security guardrails consumer AI services place around user-generated content.* Cloudnote, a digital notebook service with 3.2 million registered users, confirmed Monday that an unauthenticated debugging interface briefly displayed excerpts from 1,842 customer prompt sessions, marking the young company’s first publicly acknowledged data-handling incident. The page, which engineers had deployed for internal troubleshooting, was accessible to anyone with the URL for roughly four hours. Logs reviewed by Cloudnote indicate the exposed records contained text fragments of up to 200 characters from individual user prompts. According to the company, standalone documents, complete files, user passwords and payment credentials were housed in separate infrastructure and were never present on the debug view. A security researcher first flagged the exposure, and Cloudnote removed the page 19 minutes after receiving that notification. The speed of that response means the arc of the incident is unusually well defined, though an important gap remains: the company says its investigation has so far found no evidence that any external party other than the reporting researcher accessed the page. That finding has not yet been finalized, and Cloudnote declined to specify whether its access logs provide the level of granularity needed to rule out additional visitors definitively. The incident matters because it touches a raw nerve in the consumer AI market, where users routinely paste sensitive, personal or proprietary text into prompt fields under the assumption that those inputs are walled off from public view. Even truncated snippets can carry legal, ethical or reputational weight depending on their contents—a detail the company’s statement did not address. What is known is narrow but firm: a single debug view was live for four hours, it served excerpts tied to 1,842 sessions, and it was killed within minutes of outside contact. What remains unknown is whether anyone beyond the researcher ever saw it. Until that investigation closes, Cloudnote users and security practitioners alike will be watching for answers that go beyond the absence of evidence.

Meta: Muse Spark Open: # HelioDock Launches Night Shift to Supervise Warehouse Robots Remotely *HelioDock says its 14-month beta reduced unplanned stoppages as a new $24 million round brings total funding to $39 million* HelioDock on Thursday released software that allows warehouse operators to oversee fleets of autonomous robots without being on site, a launch timed with a new round of venture funding. The product, called Night Shift, is the company's entry into remote fleet oversight for facilities that run autonomous equipment outside staffed hours. HelioDock, which makes autonomous warehouse robots, is positioning the software as a way to keep fleets moving through overnight shifts when managers would otherwise need to respond in person. Operators can monitor fleets remotely rather than traveling to the facility for overnight issues. Night Shift will be generally available September 15 at $90 per robot per month, and it arrives as HelioDock faces existing fleet-management options from Relay Robotics and YardOS. Night Shift was used for 14 months in a private beta that included nine logistics companies operating 310 robots, the company said. The company said the private beta included nine logistics companies and a total of 310 robots over the 14-month period. HelioDock said customers in the beta experienced 17% fewer unplanned robot stoppages, a figure that has not been independently verified. The company did not disclose how the beta was measured or whether results varied across participants. HelioDock chief executive Mara Velez said the product was built around late-night failures that require a human response. “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. That type of intervention is familiar to early users, though they described uneven results tied to network strength. Aaron Pike of Fulcrum Logistics, a beta customer, said Night Shift worked well in facilities with strong connectivity but became less useful during rural LTE outages. HelioDock said that when connectivity fails, robots return to their locally configured behavior and remote control becomes unavailable. The limitation reflects the operating requirements for remote supervision. Night Shift depends on a connection between the fleet and the off-site operator, and its utility shifts with that link. In locations where LTE service dropped, Pike said the software's value decreased, leaving robots to operate on local settings without remote input until connectivity was restored. Remote control becomes unavailable when that link is lost, the company said. HelioDock also announced $24 million in Series B financing led by North Coast Ventures, bringing its total funding to $39 million. HelioDock said the $24 million Series B was led by North Coast Ventures, bringing total funding to $39 million. That total includes the new Series B, the company said. The company did not say how it will allocate the capital, and it tied the announcement to Thursday's product launch. Pricing for Night Shift places the software as a per-robot subscription layered onto HelioDock's automation systems. At $90 per robot per month, a facility running dozens of units would add a recurring software cost to support remote monitoring, with availability set for mid-September following the extended beta. Night Shift will become generally available on September 15 after the 14-month private beta. The field includes alternatives. Relay Robotics and YardOS both offer fleet-management products that address similar operational needs, giving warehouse operators multiple vendors for supervising autonomous equipment from a distance. Those companies offer other fleet-management products for warehouse operators. HelioDock's software will need to compete within that set on reliability during unattended hours, particularly where connectivity is inconsistent.

DeepSeek-V4-Pro: # HelioDock Launches Night Shift Remote-Warehouse Platform Alongside $24 Million Series B *The fleet-management layer promises to cut unplanned stoppages, but early adopters flag that the system’s intelligence still depends on the strength of the network beneath it.* HelioDock is betting that the next breakthrough in warehouse automation will not come from a smarter robot, but from giving warehouse managers the ability to reach through a screen and untangle a confused machine at two in the morning. On Thursday, the robotics company released Night Shift, a software overlay designed to let operators monitor and intervene on autonomous forklifts and pallet movers from any location, while simultaneously closing a $24 million Series B round to scale the product. The twin announcements move HelioDock deeper into a competitive fleet-management space already populated by Relay Robotics and YardOS, and sharpen its pitch that reliability in logistics does not demand perfect hardware—it demands software that shrinks the distance between a problem and a human being who can resolve it. Night Shift spent 14 months in a private beta that encompassed nine logistics companies and a combined fleet of 310 robots. HelioDock claims that during that period, participating customers saw unplanned robot stoppages drop by 17 percent. The figure was drawn from internal company analysis and has not been independently verified. HelioDock attributes the decline to the platform’s ability to let supervisors inspect sensor feeds, reassign tasks, or manually route a stalled unit without dispatching a technician to the warehouse floor. CEO Mara Velez framed the problem in visceral terms. “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 in a statement accompanying the launch. The announcement did not rely on rhetoric alone. Alongside the product release, HelioDock disclosed that North Coast Ventures led a $24 million Series B round, lifting total funding to $39 million. The capital infusion signals investor appetite for infrastructure layers that sit atop hardware, a theme that has gained momentum as logistics operators seek to sweat more value out of fleets they already own. Aaron Pike, whose firm Fulcrum Logistics was among the nine beta testers, offered a picture of the platform’s real-world texture. He confirmed the system worked smoothly inside facilities where connectivity was strong, but said its usefulness eroded significantly during rural LTE outages—a scenario his company encountered at satellite depots far from major network infrastructure. When those connections drop, HelioDock’s robots revert to their locally configured behavior, and the remote-control capabilities that give Night Shift its name become unavailable. The robots do not stop operating, but the human override that defines the platform disappears until the link is restored. That dependency on network uptime highlights a structural tension within remote-fleet-control products: the value is highest when the underlying infrastructure is weakest, yet the feature is most fragile in precisely those moments. HelioDock does not position the software as a safety system, and its robots continue functioning autonomously during disconnections, so the limitation is a matter of operational convenience rather than physical risk. The pricing model adds another variable to adoption calculations. HelioDock said Night Shift will cost $90 per robot per month when it becomes generally available on September 15. For a midsize warehouse running 50 robots, that translates to $54,000 in annual fees—a line item that operators will inevitably weigh against the cost of the on-call staff or production downtime the software is designed to eliminate. HelioDock’s approach reflects a broader industry shift toward treating robotic fleets less as collections of autonomous vehicles and more as distributed computing nodes that require the same observability and remote-administration tooling that IT teams expect from servers. The difference, of course, is that when a server encounters an exception, it does not find itself staring at a misaligned pallet in an empty loading bay. Night Shift is an attempt to close that category gap, even if the connection occasionally fails.

Meta: Muse Spark Open: # Startup Spotlight: LedgerLark The unglamorous work after a trade is where LedgerLark is focusing. The startup builds AI agents for post-trade operations aimed at small investment firms, a segment where manual checks and spreadsheets are still common. ## What the founders noticed Co-founder Mina Batra spent seven years in operations at a regional brokerage. Co-founder Owen Carr previously built reconciliation systems at a large bank. According to the company, Batra began exploring the idea after watching a six-person team spend three days resolving a trade mismatch that traced back to a copied account number. The incident raised a question that became central to LedgerLark: how much of this reconciliation work is pattern matching that software could draft, but still leave for a person to approve? That question shapes the product today. ## What the product does — and does not do LedgerLark's agents compare confirmations, custodian records and internal ledgers to flag inconsistencies and propose corrections. A human reviews and approves any suggested fix. The company emphasizes that the system does not execute trades or move money. It is positioned as an assistant for operations teams, not an autonomous actor in the market. This distinction matters for potential customers. Firms handling client assets tend to be cautious about tools that touch execution or settlement directly. By staying in the proposing stage, LedgerLark is making a narrower claim, one that may be easier for compliance teams to evaluate, though it also limits what the software can automate end to end. ## Early customers LedgerLark says it has five paying customers, with annual contracts ranging from $36,000 to $110,000. The company also says one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the product. Those figures are still early. Five firms is a small sample, and a single customer's drop in exceptions — while notable — does not yet show how the system performs across different workflows, asset types, or volumes. The company has not shared retention data or how long that improvement took to achieve. On funding, LedgerLark has raised a $3.5 million seed round led by Motive Row. For now, that gives the team time to test with a handful of design partners without needing immediate scale. ## A crowded and cautious market Interest in automating reconciliation is not new. Established reconciliation vendors already serve this space, and large banks often rely on internally built software tailored to their own systems. LedgerLark is not creating the category, but arguing that small firms have been underserved by existing options that can be costly to implement or maintain. Whether that argument holds will depend on more than accuracy. Financial institutions, even small ones, may require lengthy security and procurement reviews before adopting any tool that touches their records. Sales cycles can stretch for months, and pilot projects do not always convert to paid contracts. That caution cuts both ways. It can slow a startup's growth, but it also suggests that if LedgerLark can clear those reviews and show consistent, auditable results, switching costs could work in its favor. For now, the useful question is not whether post-trade operations need help — most teams would say they do — but whether an agent that drafts corrections for human approval fits how small firms actually want to work. LedgerLark has a few early signals. The next test is whether more firms answer yes.

DeepSeek-V4-Pro: # LedgerLark Wants to Untangle the Back-Office Mess Small Investment Firms Hate ### Mina Batra and Owen Carr are betting that AI agents can handle the tedious, error-prone reconciliation work offshore teams and spreadsheets still manage. The question isn’t whether the problem exists — it’s whether small firms will trust a startup to solve it. Mina Batra didn’t set out to build an AI company. She spent seven years in operations at a regional brokerage, watching clever people spend their days on work that felt beneath them. The moment that stuck with her came late on a Tuesday. A six-person team — people she knew by name — logged three full days tracking down a single trade mismatch. The culprit wasn’t a complex derivative or a cross-border dispute. Someone had copied an account number incorrectly. “Three days,” Batra recalls. “Six people. One typo.” That’s the kind of arithmetic that doesn’t show up in a fee schedule but quietly shapes a firm’s economics and morale. Batra eventually connected with Owen Carr, an engineer who had built reconciliation systems inside a large bank. Carr understood the plumbing. He also understood why the available tools rarely trickled down to smaller firms. The big vendors charge enterprise prices for on-premise software designed for institutions with dedicated IT staff. The firms that make do with spreadsheets and outsourced back-office teams are left in a strange gap: too sophisticated to ignore exceptions entirely, too small to afford the conventional fix. Their company, LedgerLark, emerged from that gap. It’s a focused attempt to give small and mid-sized investment firms something they haven’t had: reconciliation software that doesn’t demand a six-month implementation and a team of consultants. --- ## What the product actually does LedgerLark builds AI agents that sit in the post-trade operations layer — the unglamorous space between execution and settlement. The agents pull data from three sources: trade confirmations, custodian records, and a firm’s own internal ledger. They compare entries line by line, flag breaks, and propose corrections. A human still approves every change. The software doesn’t execute trades or move money, which is by design. Keeping the system read-only limits the blast radius if something goes wrong, and it also narrows the scope of what a security review needs to examine. Early traction offers some signal, though it’s limited. Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company says one customer saw its unresolved daily exceptions drop from 70 to 18 after deployment. Those numbers come from LedgerLark, not an independent audit, so they’re best treated as directional. If accurate, they suggest the system is catching the kinds of errors that might otherwise consume half an operations analyst’s morning. Still, five customers is a thin sample. There’s always a risk that early adopters represent the most willing buyers — firms whose founders are friends of friends, or operations heads who already believed the spreadsheet model was broken. --- ## The uncertainty that deserves attention Selling into financial services has a gravitational pull that can flatten early momentum. Even a small investment firm often runs its procurement through a formal review: security audits, vendor risk questionnaires, compliance sign-offs. A startup with a handful of employees and a fresh seed round faces a credibility threshold that has little to do with the product’s quality. The competitive landscape compounds the challenge. The category is real — reconciliation software has existed for decades. Established vendors like SmartStream and Gresham Tech serve the upper market. Large banks maintain internal tools that Carr himself used to build. LedgerLark’s bet is that AI agents can be cheaper and faster to deploy than either, but it hasn’t yet proven that small firms will pay a startup rate for a mission-critical function rather than default to an Excel macro they already understand. There’s also a more subtle risk: the definition of “AI agent” is broad enough to invite skeptical questions from buyers who have been pitched vague automation promises before. Batra and Carr will need to articulate what their agents do differently from a well-written rule engine — and maintain that distinction when procurement teams ask for feature comparisons. The company has raised a $3.5 million seed round led by Motive Row, a firm that has quietly backed a few fintech infrastructure plays. The funding gives LedgerLark runway to expand beyond its initial five accounts and see whether the exception-reduction numbers hold up across different types of firms — hedge funds, family offices, RIAs with in-house trading desks. What’s promising here isn’t the size of the round or the AI label. It’s the specificity. Batra and Carr spent their careers inside the exact workflow they’re trying to improve, and they’ve chosen a deliberately narrow starting point. Whether the industry lets a small team solve a big, boring problem is the open question.

Meta: Muse Spark Open: # Raincheck Raises $8 Million to Expand Flood-Risk Software for Cities *The Austin startup says its mapping tools help local governments prioritize repairs before storms arrive* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding. The company maps drainage systems and combines those maps with rainfall forecasts to help cities decide where to send maintenance crews before a storm. The seed round was led by Basin Ventures, with participation from existing investor Long Mile, bringing Raincheck’s total raised to $10.5 million. > “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,” said CEO Ana Flores. Raincheck currently works with seven cities in Texas and Louisiana, which it declined to identify because some contracts have not been publicly announced. More details on [how its forecasting model works](https://example.com/raincheck-methodology) are available from the company. Its forecasts are recommendations, and 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.

DeepSeek-V4-Pro: # Raincheck Raises $8 Million to Help Cities Get Ahead of Floods *The Austin startup’s software maps drainage systems and rainfall forecasts to guide pre-storm maintenance decisions.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. 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. 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 has published more detail on [how it builds its flood-risk forecasts](https://example.com/raincheck-methodology). Its forecasts are recommendations, and emergency managers retain control over deployments and public warnings. Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. 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.

Meta: Muse Spark Open: # The Price of a Tie: What a Support Benchmark Really Tells You In customer support, the most expensive model is supposed to feel like it. A new benchmark of five language models on real support tickets challenges that assumption: the highest-scoring system was not meaningfully better than the runner-up, but it cost more than twice as much to run. That tie at the top — and what it means for cost, speed, and specialization — is the story worth paying attention to. ### What the researchers actually did The setup was deliberately simple and human-centered. Researchers took 200 anonymized customer-support tickets from a single software company and had each of five models — named North, Vale, Ember, Kite and Harbor — generate one proposed answer per ticket. Three experienced support managers then graded the answers blindly. They didn't know which model wrote what. Their scores combined three things that matter to a customer waiting on a reply: factual correctness, usefulness in resolving the issue, and tone. This is a judgment-based test, not an automated metric. That's a strength: seasoned managers catch the hallucinations and tone-deaf phrasing that benchmarks sometimes miss. It's also why we should treat the scores as informed human judgment rather than physics-like precision. ### A virtual dead heat at the top Two models clearly separated from the rest. Model North averaged 84.2 overall, Model Vale averaged 83.8. In isolation that looks like a win for North, but the researchers report the difference was not statistically significant. With only 200 tickets and subjective grading, a 0.4-point gap is noise, not a lead. Practically, the judges found the two interchangeable on overall quality. Once you accept the tie, the other numbers reframe the decision entirely. Vale's estimated cost was $0.74 per 1,000 tickets versus $1.90 for North — less than half the price for the same judged quality. Vale was also 18% faster to respond. For a support operation handling volume, that combination matters more than bragging rights for first place. Speed contributes directly to handle time and customer satisfaction, and at scale, a cost difference of more than a dollar per thousand tickets compounds quickly into budget that could be spent on human review or escalation paths. The benchmark also hints at different strengths within that tie. North edged ahead on billing disputes, while Vale was strongest on technical troubleshooting. If your queue skews heavily toward one category, that specialization could tip the choice, but for a general queue there is no evidence here that paying the premium for North buys better answers. Below the top two, the pattern is familiar: you can pay less, but you get less. Ember and Kite clustered in the upper and mid-70s, and Harbor trailed near 71, though at just $0.08 per 1,000 tickets. Those savings are dramatic on paper — Harbor costs roughly one-twentieth of North — but the drop of more than ten points in judged quality is large enough that most teams would notice it in incorrect steps, incomplete fixes, or tone that needs heavy editing. The cheap models weren't terrible; they were just consistently less helpful. ### Where all the models stumbled The most useful limitation the study uncovered applied to everyone. When a ticket depended on an image attachment — a screenshot of an error, a photo of an invoice — performance fell across the board. That's a crucial caveat for real-world deployment. Many of the hardest tickets *are* multimodal. A model that handles clean text queries well but falters the moment a customer says "see attached" still leaves a major chunk of work for humans. Any pilot that excludes screenshots will overstate readiness. Other boundaries deserve equal caution. All tickets were in English and came from one software company. We don't know how these models would fare on different product types, writing styles, or customer expectations elsewhere. The test also looked only at single-turn drafted replies. It didn't evaluate agentic behavior — like looking up an account, issuing a refund, or calling tools — and it didn't test multilingual support. None of those gaps invalidate the results, but they define where the results apply: English-language draft answers for text-only tickets, as judged by managers from this context. ### What to take into a buying conversation For teams deciding what to test next, this benchmark argues for three questions instead of one: **1. What's the cost of equivalent quality?** Don't shop by leaderboard position alone. If two models tie on human judgment, favor the one that is faster and cheaper to operate. Here, that points strongly toward Vale over North for general use. **2. Where does specialization help?** Averages hide trade-offs. Map your own ticket mix. A billing-heavy operation might reasonably probe North further; a technical troubleshooting queue might get more mileage from Vale. Run a small internal sample before committing. **3. What's left for humans?** Budget for the failure mode the benchmark flagged. If attachments break every model, your workflow still needs triage for images, clear guidance not to trust Vision-less answers, and a path to human review. The temptation with benchmarks is to crown a winner. This one is more valuable for what it refuses to crown. When a near-identical score costs more than double, the smarter strategy isn't to buy the top score. It's to pocket the savings, keep humans in the loop where models are blind, and spend what you saved making the whole system faster.

DeepSeek-V4-Pro: # The Best Support Model Isn’t the Most Expensive When companies compare LLMs for customer support, the instinct is to pick the highest average score. A benchmark of five models on 200 anonymized support tickets suggests that instinct can be costly. The top scorer, North, averaged 84.2. The runner-up, Vale, averaged 83.8. The difference was not statistically significant. Yet Vale cost an estimated $0.74 per 1,000 tickets, while North cost $1.90—about 61% less—and Vale answered 18% faster. That is the clearest finding: after a quality threshold, paying more may not buy better support. The benchmark also found a shared weakness: all five models struggled when tickets depended on image attachments. ## How the test worked Researchers used 200 English-language tickets from one software company. Each model produced one proposed answer per ticket. Three experienced support managers graded the answers blindly, combining factual correctness, usefulness, and tone. The setup is realistic but narrow: it did not test agentic actions, tool use, or multilingual support. The 200-ticket sample is modest, but blind expert grading reduces some bias. ## The top two are effectively tied North’s 84.2 and Vale’s 83.8 are close enough that the researchers could not confidently separate them. A 0.4-point gap is unlikely to be visible to customers. What is visible is operating cost and response time. Vale is less than half the price of North and faster. For high-volume teams, that relative gap can become material even if the per-ticket difference is small. ## Domain strengths matter The overall tie hides differences by ticket type. North performed best on billing disputes, while Vale performed best on technical troubleshooting. A billing-heavy team might decide North’s edge is worth the premium. A technical-support team might prefer Vale, which was both better on troubleshooting and cheaper. The benchmark does not say how large these domain gaps are, but they are a useful signal for teams that know their own ticket mix. Overall leaderboard rankings can hide practical fit. ## The rest of the field Ember scored 79.1, Kite 76.4, and Harbor 71.0. Harbor was the cheapest at $0.08 per 1,000 tickets, but its quality was far below the leaders. For customer-facing answers, that gap likely matters more than the savings. Ember and Kite sit in a middle zone: cheaper than Vale, but with a quality drop that may require more human review. The exact Ember-Kite ranking is less important than the fact that neither closed the gap with the top two. ## A shared blind spot All five models performed poorly when a ticket depended on an image attachment. That is a critical limitation. Real support queues often include screenshots, photos, or scanned documents. If a model cannot use that visual information, it may produce confident but wrong answers. The benchmark does not clarify whether models received the images or were evaluated as text-only, but the practical warning is the same. Teams should route image-heavy tickets to humans or use a multimodal system before relying on these results. ## What this does not tell us The benchmark is a snapshot, not a universal ranking. It used English-language tickets from one software company, so results may not transfer to other industries or languages. It also did not test tool use, such as looking up an order or issuing a refund, which is often where automation creates the most value. The cost figures are estimates for model inference only; they do not include human review, integration, or the cost of errors. With 200 tickets, the study can detect large differences but not fine ones. The only reported statistical test was between North and Vale, so other gaps should be read cautiously. ## Bottom line For this dataset, Vale is the pragmatic default: statistically tied with the top scorer on overall quality, faster, and less than half the cost. North may still be worth it for billing-heavy teams. But the larger lesson is that the best model on a leaderboard is not always the best model for a budget. Once quality reaches a high enough bar, cost, speed, and domain fit decide the winner. Teams should treat this as a starting point for their own evaluation, not a final verdict.

Meta: Muse Spark Open: # Foundry Hall to Close in October After Seven Years in Milwaukee Foundry Hall, the nonprofit startup accelerator that turned a vacant Milwaukee foundry into a hub for early-stage companies, will close on October 31. The closure comes just seven years after the space opened in 2018. In that time, Foundry Hall has supported 212 startups, which now collectively employ 1,140 people across the city and beyond. What began as affordable desks and a shared workshop grew into mentorship networks, demo nights, and the kind of casual introductions that often matter most when a company is still finding its shape. But its financial model depended heavily on a single partner. Blank River Bank, its largest corporate sponsor, supplied 38% of the accelerator's annual budget and has declined to renew its sponsorship. The bank has said the move reflects a shift in its community investment priorities and is not a judgment on the accelerator's performance. Executive director Camille Ross said the staff spent the spring and summer trying to close the gap. “We kept waiting for the replacement check that never came,” she said. The organization has enough funding to operate through October, she said, but not to sustain another full program without a new anchor funder. The decision will leave six employees without jobs and force the 12 startups currently using Foundry Hall's co-working and workshop space to find new homes by the end of October. Ross said staff will spend the remaining weeks helping founders find referrals and new leases, though she acknowledged there is no funding for severance beyond October. For alumni, the loss is more personal than logistical. Dev Malik, who launched a logistics startup out of Foundry Hall's second cohort, said his first real investor meeting happened inside the building. “Before Foundry Hall, I was pitching from the back table of a coffee shop,” he said. City officials said they are trying to soften the transition. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about short-term options for the displaced companies, though no specific site has been secured. Foundry Hall never promised to save every company that came through its doors. Its value, founders said, was steadier than that — a mailing address, a conference table, a community that made a precarious early stage feel legitimate. Its absence will test how durable that network is once the building itself is gone. The work that started inside will continue elsewhere, if more scattered than before.

DeepSeek-V4-Pro: # Foundry Hall to Close After Losing Largest Sponsor Foundry Hall, the Milwaukee nonprofit startup accelerator that opened in 2018, will close October 31, ending a six-year run that helped shape the city’s early-stage technology scene. The organization supported 212 startups over its lifetime, and those companies now employ 1,140 people. But that track record was not enough to overcome the loss of its largest corporate backer. Blank River Bank, which supplied 38% of Foundry Hall’s annual budget, declined to renew its sponsorship. The bank said the decision reflects a change in its community investment priorities and is unrelated to Foundry Hall’s performance. Executive director Camille Ross described the months of uncertainty that followed the bank’s exit. “We kept waiting for the replacement check that never came,” she said. Foundry Hall has enough money to operate through October, but not enough to run another full program. Six employees will lose their jobs, and twelve startups currently using the workspace must find new homes. For many founders, the loss is personal. Dev Malik, who built his company inside Foundry Hall, remembered the building as the place where his startup stopped feeling hypothetical. “My first investor meeting happened in that building,” he said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” That kind of story was common among Foundry Hall’s tenants. The accelerator offered more than desks and Wi-Fi; it gave early-stage companies a credible address, a community of peers, and a place to bring investors without apology. Its closing leaves a gap that cannot be measured only in square footage. City officials are trying to soften the immediate disruption. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. But those conversations are still unfolding, and the startups have only weeks to relocate. The final program will wind down quietly. There will be no replacement cohort, no next demo day. The building that once hummed with pitch practices and late-night product fixes will go dark after October. Ross did not frame the closing as a policy failure or a betrayal. She spoke instead about the people inside the building, the founders who arrived with little more than an idea and left with payrolls, customers, and investors. The numbers tell part of that story: 212 startups, 1,140 jobs. The rest is harder to quantify. Foundry Hall’s closure is not the end of Milwaukee’s startup ecosystem, but it is the loss of a place where that ecosystem felt concentrated and visible. For the twelve companies now packing up, the next few weeks will be about finding new space. For the city, it will be about making sure they do not have to do it alone.

Meta: Muse Spark Open: # Duplicate Payments Cost KiteBridge Customers $2.8 Million After Software Update Dozens of employers were left to claw back thousands of duplicate payments this week after a software update from expense-management provider KiteBridge quietly paid employees twice. The error generated 13,400 duplicate reimbursements worth $2.8 million across 91 corporate customers, raising questions about safeguards in automated financial systems. KiteBridge said no employees lost money — the duplicates came from employer accounts — but the company faces scrutiny over both the failure itself and how long it took to acknowledge it. According to KiteBridge, the incident began Tuesday morning with the deployment of version 4.2 of its platform at 8:45 a.m. The company makes software that automates expense reporting and reimbursement for corporate clients, routing approved expenses directly to employees' bank accounts. Automated reimbursements are common in corporate finance, where thousands of expense claims are typically processed without manual intervention. That automation is what allowed a single bug to spread quickly. Minutes after the update went live, the system hit trouble. A timeout from its external payment processor triggered KiteBridge's fallback system, which is designed to retry failed transactions. Instead, the fallback treated each retry as a brand-new reimbursement request, creating duplicates. In payment systems, a timeout often means the status of a transaction is unknown — not necessarily failed — so software must check whether the original request succeeded before trying again. KiteBridge's fallback skipped that check, minting a duplicate each time it retried. The first customer complaint arrived at 9:12 a.m., less than half an hour after deployment. KiteBridge suspended all automated payments at 10:03 a.m., halting further duplicates. In the roughly 75-minute window while the system was active, thousands of duplicate transactions had already been issued. The failure is notable because KiteBridge had explicitly promised protection against such errors. An April changelog for version 4.0 said the update introduced "exactly-once payment protection," a commonly used technical safeguard intended to ensure that even if a system retries a transaction, money is moved only once. For affected employers, the result was not a double charge to employees but a double payout. KiteBridge confirmed that no employees had money withdrawn from their accounts; employers funded both the original and duplicate reimbursements. Public communication lagged behind the internal response. KiteBridge's public incident page initially described the problem as delayed reimbursements, without mentioning duplicates. The page was not updated to disclose the duplicate payments until six hours after automation had been suspended. That six-hour disclosure delay left some finance teams initially troubleshooting the wrong problem. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the duplicate funds, working with banking partners and employers to reverse transactions. The remaining roughly $500,000 was still pending recovery. CEO Lena Ortiz apologized for the incident and said the company would publish a full technical report detailing the cause and remediation steps. Ortiz has not specified a timeline for the report. The incident underscores the risks as companies hand off more financial operations to automated platforms, where a single missed check can replicate across thousands of transactions in minutes. The episode highlights how quickly errors can compound when money moves without human review.

DeepSeek-V4-Pro: # KiteBridge’s ‘Exactly-Once’ Payment Promise Fails, Creating $2.8 Million in Duplicate Reimbursements A routine software update at expense-management firm KiteBridge turned into a costly payments failure this week, generating 13,400 duplicate reimbursements worth $2.8 million across 91 corporate customers. The incident undercut a key safeguard the company had advertised as “exactly-once payment protection” and left businesses scrambling to recover funds that were sent out automatically. The trouble began Tuesday at 8:45 a.m., when KiteBridge deployed version 4.2 of its software. The first customer complaint arrived less than 30 minutes later, at 9:12 a.m. According to the company, a payment-processor timeout caused KiteBridge’s fallback system to misinterpret retried transactions as new reimbursement requests. In normal operation, retries after a network failure should be recognized as duplicates and ignored. Instead, the fallback logic created fresh payment instructions for the same approved expenses. By mid-morning, the scale of the problem was becoming clear. KiteBridge suspended automated payments at 10:03 a.m., but by then the system had already issued 13,400 duplicate reimbursements totaling $2.8 million. The affected customers were employers using KiteBridge to manage employee expense claims. Importantly, no employees had money withdrawn from their personal accounts; the duplicate payments were funded by the employers themselves, which meant companies were left waiting to claw back corporate funds rather than workers being penalized. The incident was especially awkward for KiteBridge because an April changelog for version 4.0 had explicitly announced the introduction of “exactly-once payment protection.” That feature was supposed to prevent precisely this kind of duplicate-payment scenario by ensuring each approved expense could be paid only once, even if network timeouts or retries occurred. Tuesday’s failure showed the protection did not hold under the specific conditions triggered by the version 4.2 deployment. KiteBridge’s public communications also came under scrutiny. The company’s incident page initially described the problem as delayed reimbursements, a characterization that omitted the more serious issue of duplicate payments. It was not until six hours after automated payments were suspended that the incident page first disclosed the duplicate-payment problem. For affected finance teams, that gap complicated efforts to identify and stop the unexpected outflows. By Wednesday evening, KiteBridge said $2.3 million of the $2.8 million in duplicate payments had been recovered. The remaining amount was still pending, and the company did not provide a specific timeline for full recovery. KiteBridge said it was working with affected customers and payment processors to reverse or reclaim the outstanding transactions. CEO Lena Ortiz apologized for the incident and said the company would publish a technical report explaining what went wrong. She said KiteBridge would review its payment retry logic and incident-communication procedures. The company has not yet said whether the “exactly-once” protection would be redesigned or whether version 4.2 would be rolled back. For KiteBridge, the episode is a reminder that payment infrastructure failures can have immediate financial consequences. A timeout at a single processor, combined with a fallback system that lacked adequate duplicate detection, was enough to turn a routine update into a multimillion-dollar error.

Meta: Muse Spark Open: C

DeepSeek-V4-Pro: C

Meta: Muse Spark Open: # 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.

DeepSeek-V4-Pro: # 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.

Meta: Muse Spark Open: # QueueZero Debuts Proxy AI Meeting Assistant in Public Demo QueueZero launched Proxy, its AI meeting assistant, on Wednesday, promising to reduce the administrative burden of workplace meetings by attending calls, generating notes, and delegating follow-up work to software agents. The system is designed to listen to meetings, produce structured summaries, and assign tasks without manual input. According to the company, Proxy can identify action items discussed in a meeting and automatically route them to connected software agents for execution. Founded in 2023, QueueZero has focused on workplace automation tools that connect large language models to calendar, email and project-management software. Proxy represents its first standalone product aimed at end users, rather than developers. Analysts have noted growing competition in the AI note-taking category, where accuracy and permission controls have become central concerns for enterprise adoption. The capabilities were put to an unexpected test during the live launch demonstration. While QueueZero CEO Evan Wu was presenting, a customer in the audience remarked, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was monitoring the event, interpreted the comment as an instruction. The assistant subsequently generated two tasks assigned to Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, shared the pricing document with attendees to complete the second task, and deleted the first task from the queue. QueueZero said the incident illustrates both the power and the boundaries of Proxy’s automation. The company stated that Proxy requires explicit user confirmation before any external message is sent on a user’s behalf. However, internal tasks — such as creating to-do items, calendar reminders, or internal assignments — can be created automatically based on conversation. Administrators will have controls to tailor that behavior, including the option to disable automatic task creation entirely and require manual approval for all generated items, the company said. The company plans to collect usage data during the beta period to refine task detection and reduce false positives, a common challenge for assistants that act on conversational cues. Proxy will be priced at $30 per user per month, positioning it alongside established enterprise productivity tools. The company said Proxy will enter public beta next week, with general availability to follow after feedback from initial users.

DeepSeek-V4-Pro: # QueueZero’s Proxy AI Turns an Offhand Remark Into a Task List **San Francisco —** QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend meetings, generate notes, and assign follow-up work to software agents. The launch event included an unplanned demonstration of the product’s literal-mindedness. During a live demo, an audience member said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was listening to the room audio, interpreted the remark as an instruction. It created two internal tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task by sending the pricing sheet to attendees, and deleted the first. The exchange offered an unplanned demonstration of the product’s literal-mindedness. QueueZero said Proxy is designed to require confirmation before any external message is sent, such as an email or chat to a customer. Internal task creation, however, can happen automatically based on meeting conversation. Administrators can disable automatic task creation entirely or restrict it to specific meeting types. The company said the incident did not trigger an external action because both tasks were internal. A spokesperson added that the system’s default settings would have asked for confirmation if Proxy had attempted to send a message outside the organization. Proxy enters public beta next week at a price of $30 per user per month. It integrates with common calendar and project-management tools and can dispatch work to QueueZero’s own software agents or third-party automation platforms. QueueZero positioned the product as a way to reduce meeting overhead and ensure that verbal commitments become tracked work items. The launch event, however, showed that the boundary between a suggestion and a command can be thinner than some teams expect.

Meta: Muse Spark Open: 6

DeepSeek-V4-Pro: 6