Breadwinner

The Problem

Job hunting in 2026 is an unpaid 40-hour-per-week job. Tailoring an application manually takes hours of cross-referencing job descriptions, calculating metrics, and reformatting documents.

Naturally, job seekers turned to raw consumer LLMs. But unconstrained LLMs make the problem worse: as DeepMind’s StoryScope research demonstrates, generative models systematically erase technical trade-offs, over-explain mundane concepts, and hallucinate qualifications to create tidy corporate fiction. Staring at an empty document to rewrite a CV by hand is equally broken for ADHD executive function.

Industry Recognition: XERJ Corpus Hub

In October 2026, Breadwinner was selected for inclusion in the XERJ Corpus Hub—an open index of production code solutions used by autonomous coding agents to reference battle-tested architectures instead of guessing implementations.

The hub indexed Breadwinner’s typed persistence layer (feat(contract): implement JobContract and disk persistence) as a golden reference pattern:

• Deterministic State Contracts: Autonomous agents frequently corrupt state across multi-step execution. Breadwinner enforces rigid Pydantic models (JobContract) backed by atomic disk writes, guaranteeing that application stages, tailored bullets, and telemetry never drift or desynchronize.
• FastMCP Tooling Bridge: Rather than relying on brittle system prompts, Breadwinner exposes the verified experience pantry and tailoring pipeline as typed Model Context Protocol (MCP) servers, allowing coding agents (Claude Code, Antigravity, Gemini) to interact with career data safely with explicit human-in-the-loop validation.

The Architecture

Breadwinner is a local-first Python CLI and Textual-based terminal workspace designed as an aggressive sparring partner. Instead of letting an LLM free-associate experience:

1. The Verified Pantry: All career history, production metrics, architecture wins, and constraints are stored in a local, human-verified YAML/JSON source of truth.
2. Hard Grounding Constraints: The model is locked inside a rigid box. If an accomplishment or number is not explicitly verified in the pantry, the engine hard-fails generation.
3. FastMCP Agent Protocol: Exposes verified facts and tailoring actions to local agents through typed MCP tool endpoints, preventing hallucinated tool inputs.
4. StoryScope Anti-Slop Linter: Every generated bullet point is evaluated against strict heuristic rules: zero banned corporate fluff (spearheaded, testament, delve), enforced Google XYZ formula structure, and real engineering trade-offs.
5. Deterministic Headless PDF Baking: Compiles semantic Jinja2 HTML templates into clean, single-column, ATS-parseable PDFs deterministically via headless browser printing (--no-pdf-header-footer) with strict page-budget auditing, with zero cloud dependencies.

Engineering Logs

Breadwinner’s architecture and lessons learned are documented through an ongoing series on Noggin:

• Log 01: The Cultural Diagnosis — Why tech hiring is broken, why raw LLMs produce corporate slop, and building an aggressive sparring partner for ADHD executive function.
• Log 02: The Anti-Hallucination Proof — Why a resume is an empirical technical proof, entropy math behind hallucinations, and the grounded local pantry architecture.
• Log 03: The Dogfood Incident — What happened when real beta testers fed messy LinkedIn footprints and GitHub profiles into the ingestion engine.
• Log 04 (Coming Soon): The Agentic Protocol — Exposing local verified experience to autonomous coding agents via FastMCP and typed persistence contracts.

The core engine and terminal UI are currently in closed beta testing with early users.