There is a quiet, brutal truth about selling to the U.S. government: the best company does not always win. The company that read the solicitation first wins. The one that noticed, buried in Section M, that price counted for only thirty percent and technical approach for forty. The one that started writing while everyone else was still downloading the PDF. chromie.dev, a company in Y Combinator's Summer 2026 batch, was built to give small firms that head start - and it got there by way of an unusual technical detour.
The two founders, Ananth Sankaralingam and Akshay Mistry, did not set out to build government-contracting software. They set out to make web agents reliable. Most automation forces a bad trade. Old-school robotic process automation (RPA) is dependable but brittle - it follows a script until a button moves, then it breaks. The new wave of AI agents is clever but flaky - ask one to file a form and it may hallucinate a deadline. For casual work, a wrong guess is a shrug. For a federal proposal, it is a lost bid.
Their answer was to stop choosing. chromie.dev calls its approach the deterministic stack for web agents: use the model for the "why," and enforce the "how" through deterministic tooling rather than vision-based guessing. The agent reasons about what a page means, but the steps it actually takes are scripted, logged, and repeatable. Intelligent where intelligence helps. Rigid where rigidity counts.
The distinction sounds academic until you watch an agent work. A vision-based agent looks at a screen, decides where to click, and hopes the layout has not shifted since the last run. When a portal redesigns a page or a form adds a field, the guess goes wrong quietly. chromie's version keeps the reasoning - what is this solicitation, does it fit this firm, what are its deadlines - but hands the physical steps to skills that behave the same way every time. The founders describe the result as semi-deterministic: not a fully scripted robot, not a free-roaming agent, but something deliberately in between.
The pivotFrom a browser trick to a hard market
The first thing chromie shipped was almost playful: a Chrome extension that let anyone rewrite a website with plain English - hide a distracting sidebar, change a layout, automate a repetitive click - no code required. It was a neat demonstration of the underlying engine. It was also not a business.
The engine, though, was good at something specific: reading messy web pages and doing exactly the same thing every time. That is a mundane skill until you meet an industry drowning in messy pages and rigid rules. Government procurement is that industry. Contracts live scattered across federal and state, local, and education (SLED) portals. Requirements hide inside hundred-page PDFs written in clause-speak. Miss a page limit and you are disqualified before anyone reads a word of your pitch. So chromie pointed its reliable agents at the hardest customer it could find.
There is a logic to starting somewhere unforgiving. If your agents can survive federal procurement - the acronyms, the portals that time out, the clauses that contradict each other three sections apart - they can probably survive an easier domain later. The founders name healthcare, banking, and logistics as the kinds of mission-critical work they eventually want to serve. Each shares the same demand: the automation has to be right, and it has to be able to explain why it was right. Government contracting is simply the version of that demand with the least patience for mistakes.
The productWhat chromie actually does
The pitch is plain: an all-in-one business development, capture, and proposal platform to find, qualify, and win government contracts. In practice it runs as a workflow. You build a company profile once - capabilities, NAICS codes, past performance, set-asides like 8(a) or SDVOSB, and your statements of work - and that single profile powers everything downstream.
The step people underestimate is the second one. Anyone can list open solicitations - they are public. chromie tries to show you the deal before it is public, reading agency funding patterns and the expiration dates on incumbents' contracts to forecast when a recompete will land. A help-desk contract ending in January hints at a July window. That is the difference between positioning early and scrambling late.
Capture, the third step, is where the reading actually happens. A federal solicitation is not a friendly document. The instructions that decide whether your proposal is even accepted sit in Section L; the criteria that decide whether it wins sit in Section M; the page limits, submission times, and set-aside status are scattered wherever the contracting officer felt like putting them. chromie's capture engine pulls those fields out and lays them side by side - technical forty percent, past performance thirty, price thirty, twenty-five-page limit, due at two in the afternoon Eastern. What used to be an afternoon of squinting at a PDF becomes a structured brief the team can act on immediately.
The last two steps turn analysis into a bid. chromie proposes a team - a prime, the sub-contractors whose capabilities fill the gaps, the labor categories mapped to the work - and pushes that into compliant pricing and a generated first draft. Nobody at chromie pretends that draft is ready to submit. The point is that a proposal team should start from something structured and sourced, not from a blank page and a stack of unread attachments.
The tellA replay button, not a leap of faith
If one feature explains the whole company, it is an unglamorous one: replay. Every field chromie's capture engine extracts is logged back to the exact page and clause it came from. You can replay any parse, step by step, for quality assurance or a compliance review. Most AI vendors ask you to trust the model. chromie hands you the receipts.
That is the deterministic heritage showing through. In healthcare, banking, and government - the domains the founders keep naming - "the AI was confident" is not an acceptable answer when an auditor asks why a number is what it is. An invocation log that replays from any step is. It is a boring feature that happens to be the entire value proposition.
The marketWhere chromie fits
Government contracting is a large, gated market. Deltek's GovWin, HigherGov, Unanet and a growing crowd of AI proposal tools already sell into it, and thousands of firms lean on human capture managers and BD consultants who make a living reading solicitations so their clients do not have to. chromie is not trying to replace that judgment. It is trying to hand those teams the first eighty percent - the discovery, the extraction, the audit trail - so their hours go to the part that actually wins the work.
| Approach | Strength | Weak spot |
|---|---|---|
| Classic RPA | Deterministic, auditable | Brittle - breaks when the page changes |
| General AI agent | Flexible, understands intent | Guesses; hard to audit or replay |
| Human capture team | Judgment, relationships | Slow and costly to read every PDF |
| chromie.dev | Model reasoning + scripted, replayable execution | Young; proving itself in one vertical first |
The choice of first market is itself a signal. Federal procurement is unglamorous - NAICS codes, Section L instructions, set-aside rules, past-performance tables. Founders who pick that over something shinier usually care more about the problem than the pitch. And there is a fairness angle underneath. Set-aside programs exist to route work to small, capable firms, but the paperwork keeps exactly those firms out. Lower the reading cost and you widen the door.
The peopleWho is building it
Ananth Sankaralingam, the CEO, has an engineer's resume that skips across serious institutions - software engineering at LinkedIn, AI work at Children's National Hospital, research at UMIACS at the University of Maryland. His co-founder, Akshay Mistry, is a Georgia Tech engineer. Two people, based in the New York City area, running discovery, capture, teaming and proposals on top of a web-agent stack they built themselves. It is a small team for a big, dull, important problem, which is often the right size to start.
What makes chromie interesting is not that it uses AI - everyone does now. It is the restraint. In a market full of agents that promise to do everything and log nothing, chromie's bet is that the winning feature in high-stakes work is not more autonomy. It is a system you can check. For a firm staking a quarter's revenue on a single proposal, "you can replay it" may matter more than "it's smart."
There are open questions, as there are with any two-person company a few months into a hard market. A single vertical is a good place to prove a thesis and a risky place to stay; the deterministic engine is only as valuable as the number of firms who trust it with a live bid. Government buyers move slowly, and the incumbents in capture software have relationships that go back years. chromie's answer to all of that is the same answer it gives to everything else: show your work. A tool that can trace every claim back to its source is easier to adopt when the stakes are a federal award and the reviewer is a compliance officer.
For now, chromie.dev is a small team doing an unfashionable thing carefully. It reads the documents nobody wants to read, remembers where every number came from, and tries to hand a firm its head start before the competition has finished downloading the file. Whether that grows into a platform or stays a sharp tool for one industry, the instinct underneath it - make the machine explain itself - is a useful one to watch.