Flagship case study · career workflow and evidence governance

Job Hunt OS

I built and operate Job Hunt OS across Google Sheets, Apps Script, scheduled Python and Playwright workers, multi-provider review, and document validation. It collects, verifies, scores, and packages opportunities; I review each packet and submit every application manually.

Working internal application Python · Apps Script · Playwright

Evidence and current limits

The September 3, 2026 snapshot records 40,081 canonical jobs, 640 passing tests, and zero autonomous submissions. Every application decision remains human-controlled.

Screens & Evidence

Workflow dashboard & candidate scoring

**Job Hunt OS dashboard**Synthetic job and pipeline data in a source-aligned dashboard.

01 / Problem

Manual job search created fragmented tracking and token-heavy evaluation.

Evaluating opportunities across dozens of sources required repeating context, manual deduplication, and extensive review overhead.

Job Hunt OS externalized state into canonical sheets, automated ingest and deduplication via deterministic scripts, and bounded LLM evaluation to verified resume facts.

02 / Evolution

What changed as the project grew.

01 JobSniper prompt system
System instructions with strict claim boundaries evaluated jobs inside chat sessions.
02 Workflow pipeline
Google Apps Script and scheduled workers collected postings and maintained deduplicated registries.
03 Multi-provider review
Codex, Claude, and Gemini provided structured scoring and cross-review.
04 Operated system
Daily runs, automated ledger logging, and packet generation support active search.

03 / What I built

A multi-stage job intake, evaluation, and application packaging workflow.

Canonical job store 40,000+ records tracked with URL normalization and deduplication hashing.
Claim boundary enforcement Automated checks ensure resume proposals never assert unverified claims or technologies.
Scheduled collectors Playwright and Python workers poll target platforms with recovery ledgers.
Packet generator Tailored cover letter and resume markdown files ready for human review and PDF export.

04 / Architecture

From job discovery to human-reviewed application.

Working internal application

Discovery Playwright / RSS feeds

Deduplication Canonical Sheets store

Scoring Multi-model evaluation

Packet Prep Verified fact insertion

Submission 100% human-controlled

Job Hunt OS architecture showing intake, evaluation, and human submission gates.

05 / System Boundary

AI evaluates and formats; humans submit.

Deterministic tasks

Deduplication & normalization
Date checks & expired listing pruning
Fact assertion boundary enforcement
File generation and formatting

Model reasoning

Job-to-skills alignment score
Key requirement extraction
Custom cover letter drafting within facts

Human decisions

Opportunity review & prioritization
Final packet edit and sign-off
Manual portal submission

06 / Reliability

Ledgers and idempotency prevent duplicate processing.

Ingest runs write state to recovery ledgers. If an API call fails or worker halts, subsequent runs resume without duplicates.

Every candidate packet must pass strict assertion tests against the source resume facts before presentation.

07 / Outcomes

High-throughput tracking with zero fabricated claims.

Pipeline Scale

Processed and deduplicated over 40,000 opportunities with complete audit history.

Quality Control

640 automated tests enforce prompt schema compliance and claim boundary rules.

Human Governance

Zero autonomous submissions. 100% human-verified job applications.