TalentFlow — Applicant Screening Copilot

A small applicant-tracking (ATS) app where a recruiter clicks Screen on a candidate and a headless, multi-step AI agent runs in a separate Node worker (not in the request). The agent reads the role rubric, does RAG over a hiring-guidance corpus, calibrates against past shortlisted hires, runs a (deliberately flaky) background check, and produces a structured verdict — recommendation, 0–100 score with a per-criterion breakdown, strengths, gaps, and tailored interview questions. The UI polls job progress and renders a humanized step timeline.

Architecture

Next.js UI  ──POST /api/jobs──▶  jobs table (SQLite)  ◀──claims──  Node worker
   ▲                                                                    │
   └──────────────── polls /api/jobs/:id (steps + result) ◀── writes ───┘

• Next.js (App Router) + TypeScript + Tailwind — UI, API, and job polling.
• Standalone Node worker (tsx watch worker/index.ts) — claims queued jobs and runs the agent. The API only enqueues; the agent never runs in the request.
• Vercel AI SDK v7 + OpenRouter — generateText tool-loop + generateObject verdict. Embeddings via the OpenAI client pointed at OpenRouter.
• SQLite + Drizzle ORM (better-sqlite3, WAL) — one DB file shared by the app and the worker.

All AI/embedding calls are centralized in lib/ai/*.

Setup

pnpm install
cp .env.example .env.local   # then set OPENROUTER_API_KEY
pnpm db:push                 # create the SQLite schema
pnpm db:seed                 # 8 roles, 64 candidates, rubric corpus (embeds if a key is set)

.env.local:

OPENROUTER_API_KEY=...
OPENROUTER_MODEL=openai/gpt-4o-mini
OPENROUTER_EMBEDDING_MODEL=openai/text-embedding-3-small
DATABASE_URL=file:./talentflow.db
CHAOS_FAILURE_RATE=0.25

Run (two terminals)

pnpm dev      # the Next.js app on http://localhost:3000
pnpm worker   # the screening agent worker

Open the app, click Screen on a candidate (or Screen all for this role), and watch the step timeline and result fill in. View shows an existing result without re-running the agent.

Scripts

script	what it does
pnpm dev	Next.js dev server
pnpm worker	headless screening worker
pnpm db:push	apply the Drizzle schema to SQLite
pnpm db:seed	seed roles, candidates, and the rubric corpus
pnpm db:embed	backfill rubric-chunk embeddings
pnpm build / pnpm start	production build / serve

Notes

This project ships with no telemetry or error-monitoring instrumentation — it is intentionally a clean baseline. jobs.metadata is a reserved JSON column for trace-propagation headers if distributed tracing is added later.