What is the problem your project solves?

2,171 government tenders were closing in India within 14 days when we checked this morning. Finding the handful a manufacturing MSME can win means reading them — and it can't be delegated to a junior, because judging eligibility takes someone who knows the company's entire catalog and the technical depth behind every product and service in it. So Indian Manufacturing MSMEs spend their most expensive engineering hours reading tender documents, and roughly 98% of that reading ends in "not for us".

The failure is symmetric. Read too little and you miss winnable contracts — or get disqualified on a page-12 technicality after days of bid preparation and a forfeited bid deposit. Read everything and you burn the engineers who should be building.

This is why B2G lags B2B and B2C in volume despite guaranteed demand: over ₹18 lakh crore has moved through the government e-marketplace, and a quarter of central procurement is reserved for small enterprises by mandate. The buyer shows up by law — the friction is entirely on the seller's side.

How you are solving it?

Quotra runs on one doctrine: agents extract, math reckons, Claude judges, the app presents.

The feed maintains itself. A Claude Managed Agents scheduled deployment sweeps the government portals daily: on today's run it paged through the entire closing-soon listing — 218 pages, 2,171 tenders — published the 36 matching our categories through a validated ingest API, and saved the listing text of every published tender as an evidence file, so each feed entry audits back to its source. Every run is graded by Anthropic's harness against a rubric we wrote (verbatim evidence per tender, no invented data, no computed figures); a run that breaks the rules fails.

Every tender is judged before anyone opens the app. Claude compiles each tender into a typed constraint program — one model call per tender. Two gates then run in pure code: a verbatim citation gate (any quote not present word-for-word in the tender text is rejected before it renders) and a deterministic solver that evaluates the program against the company profile into GO / NO-GO / FIXABLE with proof lines. Today's pipeline: 46 model calls produced 84 verdicts. Claude never does arithmetic — money and dates are computed in code.

Every company compounds. Each company gets its own Claude Managed Agents memory store — catalog, certifications, delivered work — mounted on every session that runs for it, so nothing about the company is ever re-taught. On-demand deep-read sessions fetch the actual tender documents and upgrade citations from listing-level to document-level.

Around that core: an eligibility matrix mapped to the company's document vault, a printable bid pack, Google sign-in with all Claude keys held server-side behind a Cloudflare Worker proxy, and a recorded transcript behind every AI answer.

Prior-work disclosure: Brainstormed and ideated with enterprises understanding their pain points and developing the idea to a solution. Built the entire product spec, execution blocks and workflows for efficient execution using Claude Code.

Built during the hackathon window: the entire next js app. All end to end integrations. the entire Claude Managed Agents layer (scheduled watcher, outcome-graded runs, per-company memory stores, on-demand deep-read sessions), the neuro-symbolic constraint engine with its verbatim citation gate and deterministic solver, the server-side LLM proxy, the auto-verdict pipeline, and the live self-updating feed.

How Did You Use Claude?

Claude is the only intelligence in the product, wired at two levels.

Messages API — Claude Opus 5 + Claude Sonnet 5. Opus 5 compiles tenders into typed constraint programs and drafts eligibility and bid-pack content; Sonnet 5 carries conversational surfaces. Everything runs on a strict-JSON contract with schema validation and one corrective retry, and every call records a transcript (prompt hash, model, duration) as provenance. In production the browser never holds a key: calls go through a session-guarded Cloudflare Worker proxy with the Anthropic key as a server secret.

Claude Managed Agents — the product's autonomous muscle:

• Agents + scheduled deployments — the tender watcher sweeps government portals daily with no server of ours; version-controlled agent YAML lives in the repo.
• Outcome rubrics — every sweep is graded by Anthropic's harness against our citation law (verbatim evidence per published tender, no invented data, no computed figures). Today's graded run came back "satisfied" — and the grader independently re-fetched an unreachable portal to verify the agent's failure report wasn't a cover story.
• Memory stores — one per company: the compounding company brain (catalog, certifications, delivered work), mounted on every session for that company; plus a separate portal-craft memory the watcher maintains between runs.
• Vaults — the ingest credential is an environment_variable credential substituted only at the network edge, scoped to our API host; the agent's sandbox never sees the secret.
• On-demand sessions — deep reads triggered from inside the product fetch the actual tender documents and feed the constraint engine document-level verbatim citations.

The design principle throughout: Claude does the language; deterministic code does the deciding. The verbatim citation gate makes a hallucinated citation structurally unrenderable, and the solver — not the model — computes every verdict, rupee figure, and deadline.

What is the deployed URL for this project?

https://quotra.xyz