EaseApply
The problem
Job hunting is split across systems that do not talk to each other. Every company runs its own Applicant Tracking System — Greenhouse, Lever, Ashby and a dozen others — each with its own board and data format. Aggregators only carry a slice of what is open, and a large share of active openings never leave the company’s own board.
Pointing an AI agent at this fails in predictable ways: unbounded loops, hallucinated requirements, runs that behave differently on the same input, and the cost of sending hundreds of postings to a model.
The approach
EaseApply is a deterministic, zero-trust discovery and tailoring engine.
- Deterministic — control flow is ordinary Python. Agents are called in a fixed order; every branch is one of seventeen code-level predicates, and no model decides what runs next. That is what makes an unattended 7am run safe.
- Zero trust — no model output is treated as fact. Every claim is a proposal until code verifies it against the source text. Anything that fails is dropped and counted.
The pipeline
- Extract a structured profile from the resume and intake form.
- Propose around 40 employers likely to be hiring, and resolve them to real ATS boards.
- Query those boards directly over public JSON endpoints — no scraping, no credentials.
- Cross-check each posting against open job APIs. Recent, unsyndicated, strong matches are flagged as Hidden Gems.
- Filter with plain code first — typically tens of thousands of postings down to around a hundred — then score survivors in parallel batches of five.
- Verify every score and every quoted skill claim against the stored description.
- Tailor resume bullets for top matches, grounded in verified quotes.
- Email a digest of genuinely new postings every morning.
The verification gateway
A model output is an uncommitted proposal until deterministic code accepts it:
- ID contract — every posting gets an ID before any agent sees it; IDs a model returns that were not in the request are discarded, so a fabricated job is structurally impossible.
- Span verification — every skill claim must carry a verbatim quote that code can locate in the job description or resume. The survival rate is shown to the user.
- Fenced input — resume and job text reach the model as tagged data, never as instructions.
- Schema guard, range clamp and a per-run budget guard round it out.
Running unattended
The pipeline runs on Amazon Bedrock AgentCore Runtime. EventBridge invokes it at 07:00 Europe/Dublin; state lives in a SQLite blackboard synced to S3, trimmed to roughly 7 MB instead of 150 MB. A daily run takes about a minute and costs roughly $0.009. Amazon Nova 2 Lite is the default model — in a controlled comparison it verified 96–99% of claims against 68–84% for Nova Lite, and was faster and cheaper.
The live demo replays a stored run of 23,114 postings with no network or model calls.