A 5-agent AI pipeline that turns one master resume into job-specific, ATS-optimized documents.
2025 · AI / Career Tech · role: owner
● status: production·11 services·~2 min read
01 — overview
ResumeAgent converts a candidate's master resume into tailored, ATS-compliant DOCX files for each job application. Five specialized AI agents run as a sequential pipeline — each with a single responsibility — while a canonical JSON resume model keeps every step structured and auditable.
02 — the problem
Qualified candidates get filtered out by Applicant Tracking Systems before a human ever reads their resume, and manually re-tailoring a resume for every posting doesn't scale.
Existing AI resume tools hand the LLM full control of the document — producing verbose, unstructured, sometimes factually wrong output with formatting that breaks in ATS parsers.
03 — architecture
A layered pipeline separates AI reasoning from rendering. The REST API feeds the pipeline service, which orchestrates five agents over a structured JSON contract; a model-agnostic LLM layer swaps providers freely; deterministic FreeMarker templates render the final DOCX.
04 — engineering decisions
01
The core design decision: the LLM never touches formatting. Agents emit structured JSON, and FreeMarker templates render it into DOCX deterministically — eliminating the formatting hallucinations that plague AI resume tools.
02
One abstraction serves OpenAI, Claude, Llama, and local Ollama models (plus OpenRouter). Providers swap via configuration, keeping token costs controllable and avoiding vendor lock-in.
03
Every resume lives as versioned JSONB in PostgreSQL — a single structured source of truth that powers version tracking, diffing between tailored variants, and auditable agent outputs.
04
RSA key-pair JWT authentication, resume version history, and Spring Security guardrails — because a career platform holds some of a user's most personal data.
Explain this project
A sequential five-agent pipeline (parser → JD analyzer → semantic matcher → rewriter → ATS optimizer) in Java 21 / Spring Boot, coordinated over a canonical JSON resume model stored as versioned JSONB in PostgreSQL. The core design decision is separating reasoning from rendering: agents emit only structured JSON, and deterministic FreeMarker templates produce the DOCX — which removes the formatting hallucinations typical of LLM resume tools. A model-agnostic LLM layer (OpenAI, Claude, Llama, Ollama, OpenRouter) swaps providers by config, and RSA key-pair JWT secures the API.
05 — features
06 — stack
The repository has the code, the commits, and the decisions in context — or reach out and I'll walk you through it.