Acuity

Site: https://acuity-phyo.duckdns.org

Initial Demo Video: https://www.youtube.com/watch?v=ULcqjiKwaZw

Final Demo Video: https://youtu.be/CkgukYmBdmM

A live technical-interview platform. A candidate solves a coding problem in a Monaco IDE with an embedded AI assistant, while an interviewer silently watches a real-time telemetry + evaluation dashboard.

Features

• Live IDE mirror. The candidate types in Monaco; the interviewer sees the same code with an amber caret tracking the candidate's cursor in real time.
• Multi-file projects. Drag-drop files or a whole folder when creating a session. The candidate gets the same tree with inline rename / delete / new-file / new-folder. Edits sync live; structural changes broadcast a refetch.
• AI assistant with stacked guardrails. Pick one or more presets (hints only, no full solutions, explain don't write, syntax only, open) — they compose, strictest wins. Replies render as markdown and see the whole project as context.
• Hallucination Injector. With probability p, a second Claude pass rewrites the reply to contain plausible flaws. The interviewer sees a flag on hallucinated turns; the candidate never does.
• Hidden tests, two run modes. Stdin mode (script reads input, prints output) and call mode (LeetCode-style function-call expression with an auto-appended harness). Output comparison is JSON-aware and whitespace-tolerant.
• Interactive terminal. The candidate runs ls, cat, run, python main.py, etc. against their file tree; output mirrors to the interviewer.
• Monitoring + replay. Tab-switch detection, large-paste flags, idle-gap detection, and a replay timeline with idle bands and event markers for tab switches, pastes, and runs.
• Waiting room. Candidates wait for explicit admit; the interviewer admits or kicks from a participants popover.
• Two-phase end-interview. The session flips to summary mode immediately; Claude generates the scorecard in the background and emits scorecard_ready when finished.
• Interviewer-chosen hallucination type. Beyond the probability, the interviewer picks what kind of flaw the injector introduces (logic / off-by-one, wrong API, edge case, inefficiency, security, or mixed) to match what the interview should test.

Stack

Layer	Technology
Frontend	Next.js (App Router, TS), TailwindCSS, Monaco Editor, react-resizable-panels, react-markdown
Backend	FastAPI (Python), native asyncio WebSockets
Orchestration	LangGraph / LangChain + Anthropic Claude
Database	Supabase Postgres (SQLAlchemy 2.0 async + Alembic)
Cache / pub-sub	Redis
Code execution	Wandbox API
Auth	Supabase Auth
Deployment	Docker Compose on a VM (Caddy HTTPS/wss) — see DEPLOY.md

Quickstart

Prereqs: Docker, uv, pnpm. You'll also need a free Supabase project (URL, anon key, service-role key, JWT secret) and an Anthropic API key.

# 1. Infra (local Postgres + Redis)
docker compose up -d

# 2. Backend
cd backend
cp .env.example .env            # fill in ANTHROPIC_API_KEY + SUPABASE_*
uv sync
uv run alembic upgrade head     # apply migrations
uv run python -m uvicorn app.main:app --reload    # http://localhost:8000  (GET /health)

# 3. Frontend (new terminal)
cd frontend
cp .env.local.example .env.local    # fill in NEXT_PUBLIC_SUPABASE_*
pnpm install
pnpm dev                        # http://localhost:3000

Then open http://localhost:3000/signup, create an interviewer account, create a session, and open the candidate invite link (/join/<CODE>) in an incognito window.

Access the app at http://localhost:3000, not a LAN IP — the frontend bakes in NEXT_PUBLIC_API_URL=http://localhost:8000 and the backend CORS allow-list is http://localhost:3000.

Demo mode (no credentials)

To try the full product without a Supabase project or Anthropic key, run in demo mode. The AI returns canned deterministic responses (including a hallucinated turn) and auth is satisfied by one-click demo identities. Only Postgres + Redis (docker compose) are needed.

docker compose up -d

cd backend
cp .env.example .env            # leave SUPABASE_*/ANTHROPIC_* as placeholders
echo "DEMO_MODE=true" >> .env
uv sync && uv run alembic upgrade head
uv run python -m uvicorn app.main:app --reload

# new terminal
cd frontend
cp .env.local.example .env.local
echo "NEXT_PUBLIC_DEMO_MODE=true" >> .env.local
pnpm install && pnpm dev

Then open http://localhost:3000/login → Enter as interviewer → create a session → copy the candidate invite link → open it in an incognito window (it auto-joins as a candidate). Admit the candidate, run the interview, and end it to see the (canned) scorecard. Never enable demo mode in production.

See CLAUDE.md for the full command reference, architecture, and conventions.