Job Agent

Job applications are multi-page forms that differ on every site. Naive automation breaks on custom dropdowns, conditional questions, SPA navigation and validation errors — and, worse, reports success when it didn’t fill anything.

Automate the tedious parts — finding, scoring, filling — while never inventing qualifications, never bypassing CAPTCHAs or login walls, and never submitting without approval.

Designed and built it end to end: architecture, agent state machine, browser automation, API, dashboard and test suite.

• TypeScript
• React
• Vite
• Express
• Server-Sent Events
• Playwright (Chromium)
• Prisma
• PostgreSQL
• Docker Compose
• Vitest
• OpenAI-compatible LLM API

AI →BACKEND →FRONTEND →DATABASES →

• —Every browser interaction goes through one InteractionEngine: it fills a field, reads it back and retries with a different strategy if the read-back disagrees.
• —Application state survives every page change and is persisted after each transition, so a server restart can resume a run.

• —Real application forms are adversarial by accident: radio groups in fieldsets, JavaScript dropdowns with no <select>, async autocompletes, fields that only render when scrolled into view, conditional questions, server-side validation and SPA navigation with no URL change.
• —An agent that clicks “Next” and gets HTTP 200 can look successful while having filled nothing.

• —Built a mock career site that implements each of those patterns on its own route, so every failure mode can be tested in isolation.
• —The agent walks one page at a time: observe → detect questions → decide answers → fill → verify by reading back → check validation → find the forward control → wait for a real page change → re-observe.
• —High-risk questions — work authorisation, visa, salary, legal attestations, demographics — are never auto-answered. CAPTCHAs and login walls pause the agent for the human. A submission only counts with positive evidence: a confirmation sentence or number.
• —Loop detection watches page and question signatures; a repeated state stops the run, captures diagnostics, attempts one recovery, then asks the human.

• —The e2e suite asserts every stage separately — job opened, analysed, questions detected, answers generated, fields filled, fields verified, page changed, resume uploaded, review reached, approval required, submitted, confirmation recorded. A run that returns 200 without filling anything fails.
• —Per-run telemetry: a screenshot, DOM, visible text and structured snapshot per step, plus a JSONL log of prompts, answers, actions and verification results.

No public demo link yet. [DEMO / REPO LINK REQUIRED]

The most important feature of an agent is knowing when to stop. Most of the engineering went into pausing, refusing and proving — not into clicking faster.VERIFY — Unverified — only in: Drafted copy — confirm voice