Inspiration
Me and my cofounder spent years watching deals the customer had already said yes to sit for weeks. The buyer was rarely the hold-up. It was a security questionnaire nobody owned, a discount waiting on finance, or a contract clause legal had accepted on the last deal that nobody wrote down. Sanjit came from Microsoft's Power BI team and saw the same pattern from the data side: the answers existed, they just lived in someone's inbox.
So we started with one observation. Big B2B deals don't die in the pitch. They die waiting on the seller's own team.
What it does
Girard is an AI agent that works for the seller on big B2B deals.
- A rep forwards the buyer's email to Girard, or copies Girard on the thread.
- Girard splits it into separate asks and gives each ask an owner.
- Anything the company already decided gets answered from that earlier decision, with the source shown next to the answer.
- Anything new goes to the person who owns it (finance, legal, security, product) by email or Slack. They decide with a one-line reply.
- Limits set once, like a discount floor or terms the company won't accept, answer requests inside those lines. Reps never see the limit itself.
- The buyer's own questionnaire comes back filled in. Anything still open says "Answer to follow."
- A person on the team sends every reply. Girard never emails a buyer on its own.
Each decision becomes policy for the next deal, so the company stops deciding the same thing twice.
How I built it
- Backend in Python and FastAPI on Render. Postgres and auth on Supabase, with row-level security on every table.
- Frontend in React and TypeScript on Vercel. Email through Resend, plus a Slack integration.
- Claude is used in four narrow places: reading what a rep is asking, writing answer prose, pulling values out of uploaded documents, and reading a company's founding message. Matching, verdicts, routing and who decides are plain deterministic code. The matching library has no network, database or model access at all.
- Every drafted answer passes a six-check gate before anyone sees it. Numbers have to match the source, citations have to point at a real ruling, comparisons have to go the right direction. If a draft fails, it retries once, then falls back to a template.
- When no confirmed rule covers a question, it escalates to the owner. Duplicate questions across deals get merged, and a working-hours clock pulls in a backup if the owner goes quiet.
- Sanjit wrote the engine with Claude Code: about 96 endpoints and over 2,000 unit tests.
Challenges I ran into
Keeping the model out of decisions. Letting an LLM answer a buyer is easy. Making sure it can't promise a discount nobody approved is hard. We ended up moving every decision into code and giving the model only writing jobs.
Answers go stale. Something true last quarter can be wrong today, so every ruling carries the name of the person who made it and a date, in an append-only record.
Email is messy. Real buyer requests arrive as threads with Excel and Word questionnaires attached, phrased differently every time, and the same question rarely uses the same words twice.
And we built this bootstrapped, which meant finding design partners before we had anything polished to show.
Accomplishments that I'm proud of
- Girard is live with design partners on real deals.
- It works where people already are: email, Slack and a web app for admins. Owners never have to learn a new tool.
- The safety rules are enforced in code and in the database, not in a prompt.
- We went from idea in June 2026 to a working product and a Delaware C corp (Girard AI, Inc.) by September.
What I learned
Sellers don't want another dashboard telling them a deal is at risk. They want the work that's blocking it done.
An agent that can say "I don't know yet" earns more trust than one that always has an answer.
And when we talk about Girard, the stall gets a reaction long before the product does. "The deal is waiting on your own team" lands with almost everyone who has sold anything big.
What's next for Girard AI
- A 30-day free trial, opening to the first 10 companies this month, on their live deals.
- Early 2027: an always-on version that reads the seller's mail and calendar, keeps a map of every big deal, and shows what's still missing before signature.
- After that: contracts and quotes drafted from approved positions, supplier portals filled for the rep, and a SOC 2 audit.
- First paying customers, and a measurement of what matters: days to signature against each customer's own past deals.
Log in or sign up for Devpost to join the conversation.