ApplyPilot
Autonomous AI Agent

1Discovery — find everything

Aggregates roles from LinkedIn, Indeed, Glassdoor, ZipRecruiter and Google Jobs via python-jobspy, extended with custom scrapers for 48 Workday employer portals and 30+ direct company career sites.

Why the direct portals matter: aggregators miss roles or list them late. Going to the source widened coverage significantly — and it is the part that required real scraper engineering rather than an API call.

2Enrichment — three-tier cascade

Listings are truncated, so every full description is pulled through a deliberately tiered strategy:

This is the design decision I am most pleased with. Routing every page through an LLM works perfectly and costs a fortune. Deterministic parsing is free but brittle. The cascade means the expensive, flexible path only runs on genuinely unrecognized pages — most resolve free at tier 1 or 2. Same result, a fraction of the cost.

3Scoring — the selectivity gate

An LLM rates every discovered role 1–10 against my actual experience and preferences. Only strong matches continue. Everything else is discarded.

This gate is the most important component in the system, and it is a deliberate constraint rather than a limitation. An automation pipeline without it is a spam cannon — useless to the employer and useless to me, since a bad-fit application wastes everyone’s time and gets filtered anyway. The system is built to apply to fewer jobs, more carefully. Quality is enforced architecturally, not by good intentions.

4Tailoring — anti-hallucination guarantee

For each role that clears the gate, the LLM rewrites my resume — reordering experience so relevant work leads, and aligning vocabulary with the posting.

The hard constraint: it cannot invent anything. Companies, dates, metrics and education are treated as immutable verified facts and pass through untouched. The model’s write access is limited to emphasis, ordering, and phrasing of things that are already true.

This was not a nice-to-have — it is the whole ballgame. A resume generator that hallucinates a skill does not just produce a bad document; it produces a document that fails a human being in an interview room and burns their credibility. Constraining the model’s authority to the subset of the problem where it cannot cause harm is, to me, the central discipline of building with LLMs.

5Cover Letter — grounded, not templated

A genuine letter per role, grounded in the specific job and company — not a template with the company name substituted in.

6Autonomous Execution — the agent

The part I learned the most from. Claude Code is connected to a Playwright browser-automation server over the Model Context Protocol (MCP). MCP is what turns a language model into an actor: it gives the model a structured interface to perceive the live page and take real actions on it.

For each application the agent opens the page, identifies which ATS it is dealing with, then completes personal details and work history, uploads the tailored resume and cover letter, and answers screening questions — across Workday, Greenhouse, Lever, and the LinkedIn/Indeed native flows.

Why an agent instead of a script: you cannot hard-code selectors for a long tail of forms you have never seen. Scripted automation breaks the moment an employer customizes their Workday tenant. A model that reads the page and reasons about it is the only approach that generalizes to the unknown — and getting that to work reliably against real, inconsistent DOMs was the core engineering challenge of the project.