The Road Does Not Care Who Called It First

Hypercubic

Hypercubic just raised a $5.3M Seed round to modernize systems most of the software world politely avoids looking at.

The round was led by CIV, with participation from Y Combinator, Afore Capital, Multimodal Ventures, Pioneer Fund, Epsilon Ventures, Unpopular Ventures, and others.

That investor list matters because Hypercubic is not building another coding assistant for clean repositories and cheerful demos. The target is the mainframe estate underneath banks, insurers, governments, airlines, and retailers, where decades of business logic live inside COBOL, terminal screens, and the memory of engineers nearing retirement.

Founders Sai Gurrapu, CEO, and Aayush Naik, CTO, are former Apple AI/ML engineers. Their thesis is brutally practical: modernizing a mainframe is not a syntax-translation exercise. Teams have to recover undocumented business rules, understand how production systems behave, rebuild the application, migrate the data, and prove that the replacement still does what the institution depends on.

Hypercubic wants to turn more of that work into repeatable software. HyperDocs explains the system. HyperTwin captures expert knowledge. Hopper lets AI agents operate inside mainframe workflows. HyperLoop focuses on transformation and verification.

That is a much harder product story than “AI writes code faster.” It is also a more valuable one. Greenfield development gets the attention, but the expensive risk sits in systems that already process payments, insurance claims, government services, airline operations, and core enterprise workflows. Those environments cannot tolerate a clever rewrite that quietly changes the underlying business logic.

SiliconANGLE reports Hypercubic is already working with a large Latin American bank and a Caribbean retailer serving more than 50M customers annually. The customers are unnamed, so treat that as company-reported traction, not an audited victory lap. The new capital is expected to deepen the agents' capabilities and expand Hypercubic's capacity to serve enterprise customers.

The bigger signal is that enterprise AI is moving backward into the stack, toward the systems nobody can casually replace and everybody still needs. If Hypercubic can make modernization more computational, verifiable, and repeatable, it is not just selling software. It is changing the economics of a services-heavy category.

Congratulations to Sai Gurrapu, Aayush Naik, and the Hypercubic team. The mainframe was never dead. It was waiting for a better interface. Which supposedly “legacy” market gets rebuilt next?

DefendEye

DefendEye just closed a major investment round, and the headline is not the undisclosed check. It is what the investor mix says about the next phase of autonomous systems: the winners will need to manufacture, navigate public procurement, and prove useful when the scene is moving faster than the people arriving at it.

NovaCapital led the round, with ff Venture Capital, Hard2beat, Sunfish Partners, and In-Q-Tel participating. That is European early-stage and deep-tech capital sitting beside a U.S. strategic investor, all backing a company trying to turn aerial situational awareness from a specialist capability into something first responders can deploy in seconds.

Founder and CEO James Buchheim built DefendEye around a blunt operational problem. Emergency teams often need an overhead view before a conventional drone can get airborne. DefendEye's answer is Overwatch, an autonomous system designed to launch quickly, detect people with onboard AI, stream live video, and recover with limited operator intervention.

The company says the capital will scale production in Kraków, Poland, and Hallandale Beach, Florida, while supporting international growth. Hardware companies do not scale on software confidence alone. Assembly, quality control, connectivity, training, support, and procurement all have to survive contact with the real world at scale and under pressure.

DefendEye has added some of that commercial plumbing. Its NOBLE distribution relationship supports agency procurement, training, and sustainment. Its EAGL Technology integration connects autonomous deployment to a broader emergency-response workflow. The product is visible, but adoption lives in the system around it.

Technical BVLOS capability also does not automatically grant legal authority to operate beyond visual line of sight. In the United States, public-safety agencies still need the applicable FAA waiver or COA provision. The buyer still has to clear policy, safety, and operating requirements.

That is why this round is more interesting than another defense-tech headline. DefendEye sits at the intersection of edge AI, connected hardware, public safety, and dual-use manufacturing, where a demonstration is only the opening argument. The test is whether the system can be reliable, supportable, affordable, and ready when the first minute matters most.

Congratulations to James Buchheim and the DefendEye team. The capital is now in place to scale the machine behind the machine. Which part becomes the hardest constraint next: manufacturing throughput, procurement cycles, or regulatory operating authority?

idler

idler just raised $9M in Seed funding to build reinforcement learning environments from real production work. Paradigm partners Alpin Yukseloglu and Frankie led the round, with Y Combinator, Long Journey Ventures, and operator angels participating.

Models do not become useful at expert work because someone hands them another polished demo. They need environments where they can attempt difficult tasks, fail in measurable ways, and improve against evidence that resembles the work people actually do.

idler is building that layer.

The San Francisco company describes itself as a frontier data research lab. Its team creates evaluations, benchmarks, datasets, and reinforcement learning environments used to train and measure advanced AI models. The point is not to make a model look smart in a controlled clip. It is to test whether it can perform when the work is long, stateful, and expensive to get wrong. That makes measurement part of the product, not paperwork.

Co-founders Ivan Chub and Nalu Concepcion, with co-founder and CTO Tony Goss, have spent the past year developing thousands of coding evaluations and environments for a leading lab. idler says models improved 8x on long-horizon coding during that period, forcing the company to keep moving its own measurement bar. In frontier AI, yesterday's hard test can become tomorrow's warm-up.

It is also moving beyond bespoke work with off-the-shelf datasets for software engineering, cybersecurity, legal, strategy and operations, and enterprise safety.

ShelfLife, idler's first public benchmark, is a digital twin built from a live retailer's operations across 9 stores, 3 commerce platforms, and multiple suppliers. It asks whether an agent can handle leadership work with judgment and precision, not merely answer a tidy question.

That is why this round matters. AI has spent years competing over model size, speed, and benchmark scores. The next fight is increasingly about the environments used to train models and the realism of the evaluations used to judge them.

Paradigm leading this Seed is a bet that the data and evaluation layer can become infrastructure rather than a consulting sidecar. The harder models become to distinguish in a demo, the more valuable rigorous measurement becomes behind the curtain.

Congratulations to Ivan Chub, Nalu Concepcion, Tony Goss, and the idler team. The company is not idling. It is building the proving ground where increasingly capable models have to show their work.

For operators building with AI, what would you trust more: a higher benchmark score, or evidence that a model can survive the workflow your business actually runs?

FSH Technologies

FSH Technologies just raised a $20M Series A to attack one of technology's most expensive old habits: making government buy slow software from vendors that earn more when the work takes longer.

The round, led by Lachy Groom with Acrew Capital co-leading, brings FSH's disclosed funding to $25M after an earlier $5M seed led by Contrary. Operator Partners and Contrary joined the Series A; the company release also names Cooley and several angels. No valuation was disclosed. The bigger number is the taxpayer bill from public servants patching forms, spreadsheets, portals, and consultants that always need another quarter.

FSH Technologies was founded in 2021 by Lilly Chen and Johnson Lin. Lilly Chen, the company's CEO, came from AI/ML engineering at Meta and built FSH around a simple belief: public-sector software should work for the people using it, not merely satisfy the procurement file. Government technology has trained citizens to expect the digital equivalent of a waiting-room chair.

The company builds configurable software for cities, school districts, state agencies, and public institutions. Its work spans Philadelphia commerce programs, Pittsburgh school food-service operations, and Buffalo library donor management. FSH says its average project affects at least 57,000 people and its software has distributed more than $15.6M in grant funding. Those company-reported figures point to the real test: whether more people can actually access a service.

FSH also reports that revenue grew inside more than 30% of existing accounts, increasing 1,340% across that group, while ARR rose sevenfold after the company activated its expansion strategy in Q2 2026. This is what outcome-based delivery is supposed to create. Build something useful, let the customer expand it, and make recurring revenue the receipt instead of the promise.

The new capital will fund product development, sales, and hiring, with FSH planning to grow from 11 employees to more than 45 across engineering, design, and client strategy. COO Jason Chen and announced CTO Rodda John join Lilly Chen as the company targets new areas including transit management, public-safety scheduling, compliance, accounting, and case management.

Congratulations to Lilly Chen, Johnson Lin, the FSH Technologies team, Lachy Groom, Acrew Capital, Contrary, Operator Partners, and the investors backing a harder kind of enterprise software. Government did not become easy. The bet is that complexity pays only when software removes it. What happens when public-sector buyers demand that standard by default?

OpenAI / Rockset

OpenAI acquired Rockset because an AI product is only as useful as the information it can retrieve when a question arrives.

Announced on June 21, 2024, the deal brought Rockset's real-time analytics database and members of its team into OpenAI. The companies did not disclose financial terms. Reuters later reported, citing sources, that OpenAI used stock in a transaction valued at a few hundred million dollars. That is reported context, not an official price tag.

The strategic logic is cleaner than most acquisition language. OpenAI said Rockset's indexing and querying technology would power retrieval infrastructure across its products. The model can be brilliant, but enterprise customers still need answers grounded in current, relevant data instead of whatever happened to be in the training set.

Rockset spent years building for that problem. Founded in 2016 by Venkat Venkataramani, Dhruba Borthakur, and Tudor Bosman, the company created a cloud-native database for low-latency SQL, search, analytics, and AI applications over fast-changing data. The founding team came out of serious data infrastructure work at Facebook, Oracle, and Google.

This deal was less about buying a feature and more about buying the machinery behind reliable AI retrieval. Rockset could ingest streaming data, index it quickly, and serve low-latency queries without forcing teams to drag every answer through a warehouse-shaped obstacle course. For OpenAI, that capability belongs close to the product core.

There is a harder edge, too. Rockset's independent service was transitioned off after the acquisition, forcing customers to migrate. The technology found a bigger home while the market lost a standalone option. Strategic fit for the buyer can still create operational pain for the customer.

Brad Lightcap, OpenAI's COO at the announcement, framed Rockset as infrastructure that turns data into actionable intelligence. Venkat Venkataramani described the move as bringing powerful retrieval to AI. Those statements land because retrieval is where enterprise AI stops being a demo and starts earning trust.

Congratulations to Venkat Venkataramani, Dhruba Borthakur, Tudor Bosman, and the Rockset team. They built infrastructure that becomes more valuable as the application layer gets faster and more ambitious.

The broader signal is worth watching: frontier AI companies are becoming infrastructure companies. Models command the headlines, but durable advantage is increasingly buried in the systems that fetch, index, permission, and serve the right data before an answer is generated.

Codio

Codio secured $5M in strategic growth financing led by long-term investor Armada Investment, and the timing matters more than the round label. The capital arrived beside a new AI Literacy offering as colleges discover that access to AI and readiness for AI are not the same thing.

A chatbot license can give a student an answer. It cannot teach when the answer is wrong, when its use should be disclosed, or who owns the decision after the prompt box goes quiet. Codio is going after judgment.

The course follows a Notice, Test, Verify, Explain and Defend method aligned with the U.S. Department of Labor's AI Literacy Framework. That turns AI literacy from a vague promise into a sequence educators can teach and students can practice. The point is not to make every graduate a model engineer. It is to make them harder to fool when AI enters the workflow.

Codio already lives close to the work. Its platform gives learners real software environments while instructors see assignments, AI interactions, and progress. Responsible AI cannot be taught as a slideshow about responsibility. Students need to use the tools, inspect the output, and defend the call.

CEO Doug Hughes frames the product around judgment. Founder and CRO Phillip Snalune has pushed Codio toward applied technical learning, while CTO Maxim Kraev built the architecture underneath it. Chairman Enis Konuk brings experience scaling Qwiklabs before Google acquired it.

Armada's role is worth reading correctly. This is not a new investor wandering into education because AI made the category fashionable. Armada is a long-term backer, and founder Daniel S. Aegerter is putting more capital behind AI fluency as a baseline workforce skill.

The University of Tampa plans to use the offering in a required course this fall, and Codio plans an Applied AI for Business Schools program next. The market test is whether AI literacy can move from a technical elective into required infrastructure across disciplines.

The $5M comes without a disclosed valuation, instrument, or detailed capital-allocation map, so there is no reason to decorate the deal with imaginary precision. The product direction is clear: Codio is expanding from teaching people how to build with technology toward teaching them how to think when technology answers back.

Congratulations to Doug Hughes, Phillip Snalune, Maxim Kraev, Enis Konuk, Codio, and Armada Investment. The opportunity is producing graduates who know when to pause, verify, and own the result.

If AI literacy becomes a graduation requirement, which institution will define the standard students are actually expected to meet?

Serif Health

Serif Health just took a significant minority investment from SEVA Growth. The interesting part is what the deal says about the business forming between federal price-transparency rules and the people who must use the data.

Hospitals and health plans now publish enormous machine-readable files full of negotiated rates. “Public” sounds useful until someone opens the file and discovers that disclosure is not the same thing as intelligence.

Serif Health built Signal for that gap. The platform cleans, validates, and structures healthcare pricing data so operators can search rates, compare networks, benchmark markets, and make decisions without treating every analysis like a forensic investigation.

The company says its data layer covers 200+ commercial payers, 5,000+ hospitals, and 350B+ rates processed each month. Those company-reported figures explain why this is a growth-equity story rather than another software company waving at a regulatory tailwind.

SEVA invested through SEVA I LP. Financial terms were not disclosed, and Serif Health did not provide a valuation or ownership percentage. The capital is expected to support sales and growth initiatives, new product functionality, and data APIs. Shalin Mehta, SEVA's founder and managing partner, is also set to join the Serif Health board.

That board seat matters. Price-transparency infrastructure sits in a strange market: the raw material is mandated, the files are messy, and the buyer still needs a trustworthy answer before a contract negotiation, network redesign, or market-entry decision. The winning product is not the one that merely stores the most rows. It is the one that turns those rows into judgment an operator can defend.

Serif Health is already pushing in that direction. Signal Ask, introduced earlier in 2026, lets users query pricing data in plain language. The product does not erase the need for source checking, but it lowers the distance between a business question and the dataset required to answer it.

Rafiq Ahmed, co-founder and CEO, and Matthew Robben, co-founder and CTO, have spent years working on the unglamorous part of transparency: making published data usable. Congratulations to Rafiq Ahmed, Matthew Robben, and the Serif Health team, and credit to Shalin Mehta and SEVA for recognizing that data access is only the opening act.

Healthcare pricing has never lacked complexity. It has lacked a clean line from disclosure to decision. If Serif Health can keep shortening that line, what other regulated data markets are ready for the same treatment?

Twin1 AI

Twin1 AI just emerged from stealth with a $20M seed round and a thesis that sounds obvious only after someone says it out loud: the most valuable knowledge inside a company is not sitting neatly in a database. It is living in people's judgment, relationships, inboxes, meetings, and the reasons behind decisions.

Bessemer Venture Partners, Tribeca Venture Partners, and Aramco Ventures co-led. EJF Ventures, F-Prime, Lakestar, Notion Capital, Orrick, and a broader syndicate joined. That is more than a capital table. It is a distribution map across legal, finance, energy, and enterprise software.

Founded in 2025 by Dr. Lewis Z. Liu, Tom Cahn, Huiting Liu, and Dr. Jonathan Budd, Twin1 pairs professionals with permission-aware digital twins designed to preserve their knowledge, context, and communication preferences. Dr. Lewis Z. Liu, the company's CEO, previously co-founded Eigen Technologies, and 3 of Twin1's 4 founders came from that business. Repeat founders are not magic, but scar tissue is still one of the better enterprise software credentials.

The product works across email, meetings, documents, Slack, Microsoft Teams, Outlook, Gmail, Google Drive, and SharePoint. Individual twins act within user-set boundaries. A governed Twin Network moves context across teams, while an enterprise MCP server gives other AI agents permission-aware access.

That distinction matters. Plenty of enterprise AI tools can generate an answer. Far fewer can explain whose knowledge informed it, whether the requester should see it, and who remains accountable when the machine starts acting. In regulated industries, capability without governance is not a product strategy. It is an audit finding waiting for a calendar invite.

Twin1 says it has spent more than a year working with Linklaters, Orrick, Dechert, Customers Bank, and Aegis Energy. It reports 30%-50% automation of communications work, while Dr. Lewis Z. Liu told Artificial Lawyer that the pipeline exceeds 400 prospects. Those are company-reported signals, not audited outcomes, but they beat theoretical demand.

The capital will expand teams in San Mateo and London, push go-to-market, and deepen the technology. The longer play is enterprise first, then lower-cost self-service. Trust is easier to scale after demanding buyers have tested it.

Congratulations to Dr. Lewis Z. Liu, Tom Cahn, Huiting Liu, Dr. Jonathan Budd, and the Twin1 AI team. Does the next durable layer in enterprise AI belong to the biggest model, or the company that knows which human context the model may use?