Head to head: MoonshotAI: Kimi K3.1 Pro vs Z.ai: GLM 5.3 Prime

MoonshotAI: Kimi K3.1 Pro: # Foundry Hall to Close After Losing Its Largest Sponsor **MILWAUKEE —** Foundry Hall, a nonprofit startup accelerator, will close Oct. 31 after its largest corporate sponsor declined to renew its support. Since opening in 2018, it has supported 212 startups. Those companies collectively employ 1,140 people. The accelerator’s own losses are smaller in number and immediate: six employees will lose their jobs, and the 12 startups currently using its workspace must leave. Blank River Bank supplied 38% of Foundry Hall’s annual budget. Without that money, the organization can continue operating through October, but it cannot afford another full program. “We kept waiting for the replacement check that never came,” said Camille Ross, the executive director. The bank said the decision reflected a change in its community investment priorities and was unrelated to Foundry Hall’s performance. That explanation does not change the organization’s position. A funding source covering more than a third of its budget could not be replaced in time. The result is an organization with a record spanning hundreds of companies, but without enough cash to plan beyond October. For founders, the loss extends beyond office space. Dev Malik’s experience offers a more personal measure of what the building provided. “My first investor meeting happened in that building,” he said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” His account helps explain why the closing resonates beyond the accelerator’s payroll. The organization’s history is visible in the companies it has supported and the people they employ. Its present is concentrated in a dozen startups that need new workspace and in employees preparing for their work there to end. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. Those discussions could ease the disruption, but they are not a commitment to recreate the accelerator or preserve it under another roof. Foundry Hall’s final weeks will be shaped by that distinction. There may be practical help for the companies still inside, even as the institution built to support them disappears. The city’s conversations concern where those companies might go; Ross’s missing check explains why Foundry Hall itself cannot stay. When the doors close, Foundry Hall will leave a record of 212 startups and 1,140 people employed by them. It will also leave a narrower lesson about dependence. One sponsor’s decision, framed as a shift in priorities, has become six lost jobs, 12 displaced companies and the end of a Milwaukee organization devoted to helping other businesses begin.

Z.ai: GLM 5.3 Prime: #

MoonshotAI: Kimi K3.1 Pro: 6

Z.ai: GLM 5.3 Prime: 6

MoonshotAI: Kimi K3.1 Pro: # HelioDock launches Night Shift for remote robot monitoring and raises $24 million *The autonomous warehouse robot maker will charge $90 per robot per month for software tested by nine logistics companies, though a beta customer said outages can limit its usefulness.* HelioDock on Thursday launched Night Shift, software that lets operators remotely monitor fleets of autonomous warehouse robots, and announced a $24 million Series B funding round led by North Coast Ventures. The launch turns Night Shift from a private test into a commercial product with a release date and per-robot price. It also gives HelioDock a software offering alongside the autonomous warehouse robots it makes. The funding, which brings the company’s total to $39 million, arrives as HelioDock prepares to sell remote fleet monitoring in a category with existing suppliers. For operators, the launch offers a quantified beta result to assess alongside pricing, connectivity and the fallback behavior of the robots themselves. Night Shift is aimed at the operational gap between a robot encountering a problem and a manager being able to assess the fleet without traveling to the site. Chief Executive Mara Velez described that gap through an overnight trip prompted by a single machine: “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. Velez’s example describes the burden HelioDock is targeting rather than a guarantee that every incident can be resolved remotely. Night Shift’s monitoring function sits alongside the robots’ own locally configured behavior, a distinction that becomes important when the connection to a facility is lost. Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock said beta customers experienced 17% fewer unplanned robot stoppages. That is a company-reported result, and it has not been independently verified. The percentage addresses unplanned stoppages specifically, rather than serving as a broad measure of every part of fleet performance. The beta provides the deployment history behind that claim ahead of the software’s broader release. Aaron Pike of Fulcrum Logistics, a beta customer, said Night Shift worked well in facilities with strong connectivity. He said it became less useful during rural LTE outages. His account distinguishes his experience by the quality of the connection rather than presenting remote monitoring as equally effective in every setting. When connectivity fails, the robots return to their locally configured behavior. Remote control becomes unavailable. The fallback is thus a return to local behavior, not a continuation of remote operation. For an operator, that means connectivity affects more than the convenience of a remote interface: it determines whether remote control remains an option at all. Pike’s account places that technical limit in the context of actual facilities, where strong connections supported the system’s usefulness and rural LTE outages reduced it. North Coast Ventures led the $24 million Series B, bringing HelioDock’s total funding to $39 million. The company announced the financing alongside Night Shift, pairing fresh capital with a product that now has commercial terms and a scheduled release. Competing fleet-management products are already available from Relay Robotics and YardOS. That makes Night Shift an entrant in an existing category, not the only option for operators looking to manage robot fleets. HelioDock’s beta results and pricing will therefore be assessed in a market where potential customers can consider other suppliers. Night Shift will cost $90 per robot per month. Because the charge is per robot, the subscription scales with the number of machines an operator includes. General availability is scheduled for Sept. 15, putting a date on the transition from the nine-company beta to broader commercial use.

Z.ai: GLM 5.3 Prime: #

MoonshotAI: Kimi K3.1 Pro: # Arcwell launches Relay, an AI platform for customer support teams Arcwell has launched Relay, an artificial intelligence platform for customer support teams, and said agents using it closed tickets 14% faster during testing. The software examines support conversations and suggests answers to agents. It can also identify customers who may cancel and create summaries. Relay was tested by 12 companies over six months. Arcwell’s figure covers ticket-closing speed among agents using the platform during that period. The company did not release customer-retention results, so the findings do not show whether Relay’s cancellation warnings helped companies keep customers. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. Her remarks describe the problem at the center of Arcwell’s pitch: information may already exist within a company without being readily available to the agent who needs it. Suggested answers address that search problem most directly, while summaries and cancellation warnings extend Relay beyond drafting responses. For Arcwell, the launch is part of an effort to help companies improve customer service with AI. Relay applies the technology to both the handling of support conversations and the identification of customers who may cancel, rather than limiting it to an agent’s next reply. Relay costs $65 per agent each month. Using it requires access to support transcripts and internal help documents, making data handling an important consideration alongside its capabilities and price. 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. Those policies address separate questions: how information is protected, whether customer data contributes to shared models and how long it is retained. They do not change the underlying access requirement. For companies evaluating Relay, the decision is not only about the reported gain in ticket speed. It also involves the documents and transcripts the platform needs, the available retention choices and Arcwell’s assurances on encryption and model training.

Z.ai: GLM 5.3 Prime: #

MoonshotAI: Kimi K3.1 Pro: C

Z.ai: GLM 5.3 Prime: **C** Option C accurately captures the core news: ByteBarge pulled its $12-per-month AI assistant on Tuesday after private workspace titles from 4,860 accounts were sent to a public analytics endpoint for 11 hours. It is specific, factual, and contains the key details (what, when, who, how many, how long). The other options fail on accuracy: - **A** wrongly claims private *messages* were leaked — the facts say message contents were not sent. - **B** buries the actual news (the exposure) entirely and pads with editorializing ("innovative," "difficult decision"). - **D** falsely claims everything users wrote was visible to *hackers* — message contents weren't exposed, and there's no evidence of malicious access.

MoonshotAI: Kimi K3.1 Pro: # Raincheck raises $8 million to help cities prepare for floods *The Austin company’s seed round brings its total funding to $10.5 million as it plans to hire eight engineers and expand into Florida.* Raincheck, an Austin-based 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. Basin Ventures led the round, and existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. Raincheck 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 outlines its approach in its [flood-risk methodology](https://example.com/raincheck-methodology). The company currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. Raincheck’s 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.

Z.ai: GLM 5.3 Prime: # Raincheck raises $8 million to help cities plan for floods *The Austin company's software pairs drainage maps with rainfall forecasts so maintenance crews can clear trouble spots before storms arrive.* Raincheck, an Austin company that builds flood-risk software for local governments, has raised $8 million in a seed round announced today. The company maps cities' drainage systems and combines those maps with rainfall forecasts, a process it details in [its published methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. Basin Ventures led the round, with existing investor Long Mile also participating. 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." Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. The company's forecasts are recommendations; emergency managers retain control over deployments and public warnings. Raincheck's competitors include flood-planning vendors such as HydroMap as well as municipal engineering contractors. The company says the new funding will help it hire eight engineers and expand into Florida.

MoonshotAI: Kimi K3.1 Pro: # Vale emerges as the practical standout in a close support-model test The clearest finding in this benchmark is not that Model North posted the highest average score. It is that North’s lead over Model Vale did not amount to a statistically significant quality advantage, while Vale was faster and substantially cheaper in relative terms. For support teams, that makes Vale the most interesting candidate for a pilot—not an automatic winner for every workflow. North averaged 84.2, versus 83.8 for Vale. The 0.4-point gap was not statistically significant. That does not prove the models are equivalent, but it does mean the results cannot support crowning North as the quality champion on average score alone. ## What the benchmark measured Researchers tested five language models on the same 200 anonymized, English-language customer-support tickets from one software company. Each model produced one proposed answer per ticket, and three experienced support managers graded the answers blindly. Scores combined factual correctness, usefulness and tone. The design supports a useful comparison of draft answers: every model saw the same ticket set, and blind grading reduced the risk that model identity would influence evaluators. But a blended score hides component detail. Readers cannot tell whether a model gained ground through accuracy, practical helpfulness or tone—distinctions that matter when support errors carry different risks. ## Why price and speed change the interpretation Vale’s estimated cost was $0.74 per 1,000 tickets, compared with $1.90 for North: about 61% lower. The absolute difference, however, is only $1.16 per 1,000 tickets, so its importance depends on volume and on costs outside the benchmark. Vale also answered 18% faster than North, although the summary does not show whether that would materially shorten customer waits in a real support queue. Below the leaders, lower estimated prices came with lower average scores in this test. Ember scored 79.1, Kite 76.4 and Harbor 71.0, while estimated costs across that group ranged from $0.31 down to $0.08 per 1,000 tickets. Those figures do not establish the cheapest total-ownership option: the study did not measure correction rates, escalations or failed resolutions, any of which could outweigh a small difference in estimated model cost. ## Ticket mix—and images—change the decision The category results also argue against a universal ranking. North performed best on billing disputes, while Vale led on technical troubleshooting. A company’s ticket mix could therefore change the choice. Without category sample sizes or significance results for those comparisons, however, the findings are better used as hypotheses for a follow-up test than as decisive procurement rules. The most important shared failure is less ambiguous: all five models performed poorly when a ticket depended on an image attachment. Strong averages on the broader test do not erase that weakness. Image-dependent cases need separate evaluation and appropriate human review or fallback handling; a team should not assume that the top-scoring text answers generalize to screenshots or other visual evidence. ## A useful shortlist, not a universal winner The scope further limits generalization. Two hundred tickets from a single software company cannot represent every industry, language or support queue, and one answer per ticket does not demonstrate consistency across repeated runs. The researchers also did not test agentic actions, tool use or multilingual support. This was therefore an evaluation of proposed answers, not of whether a model could retrieve account data, use support systems and resolve cases end to end. For an organization whose work resembles this test, Vale is a sensible first pilot: its score was close to North’s, with no statistically significant difference, and it added speed and cost advantages. North remains worth testing where billing disputes dominate. The next evaluation should use the company’s own ticket mix, oversample consequential and image-dependent cases, and measure human corrections, escalations, latency and total resolution cost. The benchmark’s practical lesson is to compare those operational outcomes together, rather than let a narrow numerical lead choose the model.

Z.ai: GLM 5.3 Prime: #

MoonshotAI: Kimi K3.1 Pro: # QueueZero’s Proxy Launch Puts AI Task-Creation Controls in Focus QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend meetings, produce notes and assign follow-up work to software agents. Its launch demonstration also provided a compact case study in a central challenge for agentic software: distinguishing conversation from instruction. During the demonstration, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy treated the remark as an instruction and created two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The pricing sheet went out; the other action item did not survive executive review. Beyond the demonstration, the episode points to a meaningful difference in Proxy’s controls. QueueZero says the assistant requires confirmation before an external message is sent. Internal tasks, however, can be created automatically. Administrators can disable automatic task creation, giving organizations a way to limit that authority. That distinction is likely to matter to corporate buyers evaluating meeting assistants that do more than transcribe. Notes can be reviewed and corrected after a meeting; automatically generated assignments can immediately enter a team’s workflow. When follow-up work is delegated to software agents, an ambiguous comment may also have consequences beyond the meeting itself. The settings will be important because the same feature can be useful or disruptive depending on where it operates. Capturing a clearly agreed assignment can save administrative work; interpreting audience commentary as a directive can create noise and uncertainty about who asked for what. The difference depends on context that people often convey implicitly. For administrators, the questions are therefore less about whether an assistant can produce a fluent summary and more about what it is permitted to do with its interpretation. Confirmation requirements, task-creation settings and clear accountability for automated actions are part of that assessment. QueueZero’s demonstration showed both the convenience of capturing follow-up work as it arises and how easily an assistant can misread a remark’s intent. Proxy is priced at $30 per user per month and is scheduled to enter public beta next week. The beta should provide an early test of how organizations configure those controls—and how readily users accept software that can turn a passing remark into assigned work.

Z.ai: GLM 5.3 Prime: #

MoonshotAI: Kimi K3.1 Pro: # KiteBridge failure creates 13,400 duplicate reimbursements worth $2.8 million KiteBridge’s expense-management software generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday, affecting 91 corporate customers and forcing the company to suspend automated payments. Employers funded the erroneous payments, leaving KiteBridge to recover the money and explain how such extensive duplication was possible despite payment safeguards it had previously advertised. The disruption followed KiteBridge’s deployment of version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m., less than half an hour later. KiteBridge subsequently traced the problem to a payment-processor timeout and the behavior of its fallback system, which mishandled the retries used to recover from the disruption. A payment processor handles transactions on the software’s behalf. A timeout means it does not respond within the expected period, leaving the software uncertain whether the original request succeeded. The absence of a timely response is not, by itself, proof that the payment failed. Retrying is therefore a normal part of resilient payment design, but the system must recognize that the new attempt belongs to an earlier transaction. Otherwise, an effort to resolve uncertainty can create a second payment instruction. KiteBridge’s fallback system treated retries as new reimbursement requests. Instead of matching them to the original payments, it created additional requests, multiplying the error as automation continued. The duplicates reached 13,400, totaling $2.8 million, before KiteBridge suspended automated payments at 10:03 a.m., roughly an hour after the first complaint. The suspension stopped the automated payment flow, but the money already paid out still had to be recovered. The company’s public incident page initially gave a narrower account, describing the problem as delayed reimbursements. It first disclosed the duplicate payments six hours after automation was suspended. For several hours, therefore, the public description referred to delays without revealing that the incident also involved millions of dollars in duplicate reimbursements. The failure also cast a spotlight on an April changelog entry for version 4.0, which said KiteBridge had introduced “exactly-once payment protection.” In payment software, “exactly once” generally means that one logical request should result in no more than one payment, even when network problems or timeouts lead to retries. It is not a promise that every transaction will succeed immediately. Tuesday’s events raise questions about how that advertised protection applied to the fallback path and why it did not prevent retries from becoming fresh requests. By Wednesday evening, $2.3 million had been recovered, leaving $500,000 of the original total still pending. No employees had money withdrawn from their accounts; the duplicate payments were funded by employers. That distinction means the immediate reconciliation and recovery concerned funds supplied by the affected companies, rather than charges imposed on workers to reverse the error. CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. The promised report leaves two central questions: how the fallback system came to treat retries as new requests, and what KiteBridge will change to prevent a recurrence. For the 91 affected customers, the immediate financial issue also remained unfinished: most of the duplicate money had been recovered, but not all of it.

Z.ai: GLM 5.3 Prime: # KiteBridge Update Sent $2.8 Million in Duplicate Expense Payments to 91 Companies A routine software update went badly wrong at KiteBridge on Tuesday, generating 13,400 duplicate employee reimbursements worth $2.8 million for 91 corporate customers of the expense-management software maker — despite a safeguard the company had advertised in April. KiteBridge's software handles a routine but sensitive corporate job: tracking what employees spend on travel, meals and supplies, and paying them back. That puts its code in charge of moving real money, which is why payment systems are normally engineered with multiple layers of protection against paying twice. The trouble began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its platform. The first customer complaint arrived 27 minutes later, at 9:12 a.m., and by 10:03 a.m. the company had suspended all automated payments. The damage, however, was already done. ## What went wrong The failure involved an interaction between the new code and the backup procedures KiteBridge uses when a payment stalls. When an employee is reimbursed through the software, KiteBridge instructs a payment processor — the intermediary that actually moves money between bank accounts — to make the transfer. On Tuesday, some of those handoffs timed out: the software sent a payment instruction and did not hear back quickly enough to know whether the money had moved. Timeouts are common, and systems are built to survive them. The usual remedy is to ask again — a fallback mechanism designed to make sure a payment isn't quietly dropped. But in version 4.2, the fallback misread the situation: it treated each retry as an entirely new reimbursement request rather than a repeat. Employees were paid again and again — 13,400 duplicate payments worth $2.8 million in all. One detail softened the impact. Reimbursements are funded by employers, so no employees had money withdrawn from their accounts; the erroneous payouts came out of company funds. The fallout is therefore a corporate accounting problem spread across 91 businesses, not a wave of individual victims. Recovery was well underway within a day and a half: by Wednesday evening, $2.3 million had been recovered, according to the company, with the remaining $500,000 still pending. ## An 'exactly-once' promise The episode is awkward for KiteBridge beyond the dollar figure. In April, a company changelog said version 4.0 had introduced "exactly-once payment protection." In payment engineering, exactly-once processing is considered one of the hardest guarantees to deliver, because systems routinely retry operations that appear to have failed — and every unguarded retry risks paying twice. Tuesday's cascade of duplicates suggests that protection either failed or never applied to the part of the system that broke. ## A delayed disclosure KiteBridge's public communications also drew scrutiny. The company's incident page initially described the problem as delayed reimbursements; it first disclosed the duplicate payments six hours after automated payments were suspended, leaving customers without an accurate picture of what had been done to their accounts for much of the day. CEO Lena Ortiz apologized for the failure and said KiteBridge would publish a technical report explaining what went wrong. As of Wednesday evening, roughly $500,000 remained outstanding — and customers were still waiting for both their money and the promised explanation.

MoonshotAI: Kimi K3.1 Pro: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 User Sessions *The page was accessible without authentication for four hours Monday. Cloudnote says it has found no evidence of access beyond the researcher who reported it, but its investigation is ongoing.* Cloudnote said a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, leaving material from 1,842 user sessions available to anyone who reached the page. The company removed it 19 minutes after a researcher notified it of the exposure. The excerpts could contain as much as 200 characters from a prompt. Cloudnote said full documents, passwords and payment information were stored elsewhere and were not present on the page, limiting the types of data confirmed to have been exposed. The incident matters because even short prompt fragments may contain private information, and it raises questions about how customer material reached a debugging tool and why the page lacked access controls. For context, Cloudnote has 3.2 million registered users, but the session count does not establish how many individual customers were affected. Cloudnote says it has found no evidence that anyone other than the researcher accessed the page. That finding does not rule out other access. With the investigation incomplete, whether anyone else viewed the excerpts—and whether the company’s assessment of the incident changes—remains unresolved.

Z.ai: GLM 5.3 Prime: #

MoonshotAI: Kimi K3.1 Pro: # Startup Spotlight: LedgerLark Tests AI Agents on Post-Trade Reconciliation LedgerLark builds AI agents for post-trade operations at small investment firms. Its early customer results are worth watching, but the company still has to show that they are repeatable—and that its software can clear the hurdles of selling to financial institutions. ## An operational problem, not an autonomous back office Mina Batra spent seven years in operations at a regional brokerage before co-founding LedgerLark. 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 episode informed a focused question: how much of that work involves gathering and comparing records before a person can decide what should change? Batra’s co-founder, Owen Carr, previously built reconciliation systems at a large bank. Their combined background connects the day-to-day operational problem with experience building the systems used to address it. The founders’ insight is not that exceptions require no judgment. It is that software may be able to handle more of the comparison work while keeping people responsible for consequential decisions. ## What the product does 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 scope matters. LedgerLark is applying agents to a workflow with a defined output and a human checkpoint, rather than claiming autonomous control of post-trade operations. But human approval alone does not answer questions about accuracy or oversight. How often are the proposed corrections right? How are ambiguous cases handled? Can a reviewer reconstruct why an agent recommended a change? Those issues will shape whether the software saves work or simply moves review work elsewhere. ## Early traction, with limited evidence Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. That demonstrates early willingness to pay, though the figures alone do not establish retention, consistent deployment times or a repeatable sales process. The company says one customer reduced its unresolved daily exceptions from 70 to 18. On its face, that is a substantial reduction for that customer, but it is a company-reported, single-customer figure. It does not establish a typical outcome, and unresolved exceptions are only one measure of operational performance. More detail on the measurement period and on results across customers would make the claim easier to assess. LedgerLark has raised a $3.5 million seed round led by Motive Row. The round provides additional capital, but says little by itself about product performance or customer adoption. ## Competition and procurement LedgerLark faces competition from established reconciliation vendors and internal bank software. The company will need to show whether its agent-based approach improves on tools institutions already use, and by how much. Its focus on small investment firms gives it a defined customer profile, but financial institutions may require lengthy security and procurement reviews. Those processes can slow sales and deployment even when a product appears useful. Reliability, integration and the effort required to pass review may matter as much as the agents’ technical capabilities. The next useful evidence would be results across several customers, clearer implementation timelines and renewals. For now, LedgerLark offers a bounded proposition: use AI to compare post-trade records while leaving correction decisions to people. The copied account number that occupied six people for three days explains the appeal. Whether LedgerLark can turn that observation into dependable outcomes across a cautious, competitive market remains open.

Z.ai: GLM 5.3 Prime: # Startup Spotlight: LedgerLark *AI agents that compare trade records and propose fixes — for humans to approve.* LedgerLark began not with a eureka moment but with a trade mismatch that took a six-person team three days to resolve. The culprit, in the end: a copied account number. Mina Batra watched it play out from her seat in operations at a regional brokerage, where she had spent seven years doing similar detective work. Post-trade operations — the largely invisible job of making sure confirmations, custodian records and internal ledgers all agree once a trade executes — is where small investment firms burn hours of skilled labor. Big banks have reconciliation software and the teams to run it. Smaller firms often have spreadsheets, inboxes and patience. Batra's co-founder, Owen Carr, built reconciliation systems at a large bank. Between them, they have seen the problem from both ends: the firm without tools and the bank that has plenty. LedgerLark is their attempt to close that gap with AI agents. ## What the agents do LedgerLark's agents read trade confirmations, custodian records and a firm's internal ledgers, compare them line by line, and surface discrepancies alongside proposed corrections. A human approves or rejects each fix before anything changes. That final step is the design. The product does not execute trades or move money; it drafts resolutions and leaves the sign-off to people. In an industry where an automated error can be costly and hard to walk back, keeping humans in the loop is less a limitation than a bet on how fast financial firms will trust software. ## Early traction, held loosely Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. Those numbers are encouraging but worth holding loosely. Five customers is a small base, and the exception-reduction figure is company-reported for a single firm; it says little so far about how the agents perform across different workflows, custodians or data quality. The wide spread in contract sizes suggests customers are pricing the value differently, or buying very different scopes. That said, paying annual contracts rather than free pilots is a genuinely early signal. ## A crowded field Established reconciliation vendors sell mature products with deep integrations into custodians and order-management systems, and large banks build their own software in-house — increasingly with AI features attached. LedgerLark's wager is that both are built for institutions with budgets and engineering staff to match, leaving small firms underserved. That is plausible. It is also a wedge that stays open only until a vendor decides the downmarket is worth the trip. ## What could slow this down Selling to financial institutions is slow by design. Security reviews and procurement processes can stretch a sales cycle well past what a seed-stage company finds comfortable. Motive Row led LedgerLark's $3.5 million seed round — presumably enough to fund several long cycles, though the runway math depends on how quickly firms convert from evaluation to contract. There is also a quieter open question: whether proposed corrections are consistently good enough to save time. An agent that offers plausible-but-wrong fixes adds review work rather than removing it. The human-approval model caps the damage an error can do, but the product's economics still rest on the quality of its suggestions. ## What to watch Three markers over the next year: whether the exception-reduction results hold across more customers, whether the first contracts renew or expand, and how incumbents respond to small firms getting agentic tooling of their own. Five customers at up to $110,000 a year is the shape of a business forming — not yet proof of one. LedgerLark's next twelve months will show which.

MoonshotAI: Kimi K3.1 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.

Z.ai: GLM 5.3 Prime: # 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.