Tough Crowd
Feedback from 20 users, before you have 20 users.
- Agentic orchestration
- NLP & information extraction
- Data science & evaluation
- Cloud infrastructure
Drop in a URL. Synthetic personas research your market, then actually use your product in live browser sessions, click by click, and return a scorecard, friction findings pinned to the step that produced them, and a ranked fix-list.
“Synthetic data is a hypothesis, not a finding. The notice is non-suppressible by design.”
The problem
Pre-product-market-fit research is too slow and too expensive, so founders ship on instinct and find out why users churned after they churned.
The approach
Synthetic usage, not synthetic interviews. The personas do not answer questionnaires — they bounce off your signup form and report where. The sampling frame mandates abandoners, skeptics and critics alongside adopters, so the happy path cannot dominate by construction.
How it works
- A worker pipeline: personas, snapshot, live browser session, diffing, evaluation, findings, synthesis, artifacts
- Claude-vision Chromium sessions driving the real product rather than a recording of it
- Findings pinned to the step that produced them, with screenshots
- Re-run diffing, so a shipped fix shows up as a scorecard delta
- Built on voice-of-agents, pinned by commit and never forked
Stack
FastAPI on AWS Lambda with an ECS worker; React 18, Vite, TypeScript and Tailwind; Supabase Postgres with row-level security throughout; Stripe; PostHog; Terraform.
Where it stands
- Roughly 101 test files, with CI, deploy workflows and infrastructure as code