Job Hunt Signal

Turn one résumé into a ranked, evidence-backed job pipeline and a five-person referral plan.

Comment

Inspiration

Job searching is rarely one hard task. It is dozens of repeated small tasks: finding fresh roles, checking eligibility, comparing a job description with a résumé, deciding where to spend time, rewriting evidence, finding people who might respond, drafting messages, and remembering every follow-up. A spreadsheet stores rows but does not do that work; a generic chatbot forgets the state and can invent details.

We wanted a private job-search workspace that carries verified context from the résumé all the way to the application and outreach decision. It should reduce effort without taking away the user's control or pretending uncertainty is confidence.

What it does

Job Hunt Signal turns an uploaded résumé into a durable, owner-scoped job pipeline:

  • It safely parses PDF, DOCX, and TXT résumés, extracts high-confidence profile fields and skills, and converts exact résumé bullets into grounded evidence.
  • It scans curated first-party company and ATS sources, normalizes and deduplicates openings, rejects stale or ineligible roles, and ranks the whole Today inbox by explainable fit before pagination.
  • Each role explains matching evidence, uncertainty, and gaps. Pursuing one creates a dossier with a why-fit story, requirement coverage, tailored résumé changes, application answers, and interview preparation grounded in the exact résumé and posting versions.
  • It researches up to five appropriate referral leads, preserves source evidence and honest shortfalls, and prepares a distinct message for each person instead of one generic template.
  • It tracks manual sends, follow-ups, application milestones, interviews, corrections, outcomes, overdue work, and weekly funnel learning.
  • Normal signup gives each person a separate encrypted workspace with export and deletion controls.

The product never auto-submits an employer form or auto-sends outreach. It never fills an evidence gap with an invented claim or person.

How we built it

The product uses a Next.js 16/React 19 frontend, a FastAPI backend, PostgreSQL 16, Alembic migrations, encrypted owner-scoped repositories, and a durable worker queue. Search adapters read public first-party career sources. Matching, evidence grounding, practical application materials, and default outreach drafts are deterministic and provider-free. SerpAPI can optionally power live public-profile discovery; the older agent experiment uses Google ADK/Gemini and Phoenix, but neither is required for the practical free workflow.

We built the extension in small vertical slices so a single product decision was reflected in the database constraint, repository transaction, API schema, generated TypeScript contract, interface, and regression coverage together.

How Codex and GPT-5.6 accelerated the work

Codex was the implementation collaborator throughout the Build Week extension. It helped translate direct product feedback into small commits across the FastAPI/Postgres backend and Next.js frontend, inspect failures at their real boundary, and keep the API contract and tests synchronized with each change.

Concrete areas include:

  • owner-scoped models, Alembic migrations, transactional repositories, and generated OpenAPI/TypeScript contracts;
  • the automatic fit engine and persisted Today ordering;
  • the five-person source-backed contact bench and per-person outreach drafts;
  • evidence-pinned application materials and interview story starters;
  • normal multi-user authentication, encrypted privacy controls, and safe deployment across Vercel, Render, and Neon Postgres; and
  • the bounded PDF/DOCX/TXT parser, atomic résumé import, and upload-first UI.

During the GPT-5.6 submission pass, a GPT-5.6 Codex agent read the public API contracts and implemented the dependency-free demo workspace seeder and its sample-data guide. GPT-5.6 was used through Codex as a development model; it is not a production inference dependency and does not receive users' résumés.

Vasu and Arpita made the key product decisions: serve real job seekers rather than ship a one-run demo; rank roles by fit; find five appropriate people rather than one; make résumé upload the default; replace a shared passkey with normal accounts; prefer free first-party sources; minimize repetitive approvals; never invent evidence; and leave sending and applying to the user. Codex explored and implemented solutions under those constraints.

What was new during Build Week

The repository existed before the event as a narrower agent demo. Commit 31043a9 is the final pre-submission-period baseline. The Build Week extension starts at d1b64b1; the dated work is visible at:

https://github.com/vasubhrdwj/agentic-job-hunt/compare/31043a9...main

That extension added the durable opportunity radar, fit-ranked Today inbox, application dossiers, five-contact outreach waves, application materials, interview preparation, outcome learning, multi-user accounts, privacy controls, free-tier scanning, and secure upload-first résumé onboarding.

Challenges we ran into

Useful automation without fabricated confidence. A fit percentage looked simple but hid missing job-description and candidate facts. We replaced it with categorical, explainable assessment: eligibility, fit band, confidence, supporting evidence, and explicit missing inputs.

A referral feature that is not spam. Finding one person was too fragile; returning arbitrary names was worse. We built a diverse bench of up to five, pinned every result to public evidence, drafted separately for each person, added cooldowns and volume limits, and kept sending manual.

Turning a demo into a real account-based product. The original shared owner-key model blocked normal use. We migrated to email/password accounts and owner-scoped data without losing the existing workspace, then hardened sessions, CORS, upload body limits, migrations, exports, and deletion.

Free-tier deployment boundaries. Render sleep, worker availability, database migrations, and Vercel proxy failures can look like product failures. We separated web readiness from provider capability, added durable retries and an embedded free-tier scan worker, and made degraded capabilities visible.

Accomplishments we are proud of

  • The app now evaluates and orders opportunities before asking the user to do work, instead of recreating a spreadsheet with extra approval buttons.
  • One résumé upload creates reusable profile context and grounded evidence while preserving encryption, account isolation, and immutable history.
  • Every pursued role connects the same evidence to fit, application materials, five possible referral paths, outreach, interview preparation, and outcomes.
  • Missing evidence remains a visible product state. A shorter honest result is considered better than a polished hallucination.
  • The repository includes a free, deterministic, synthetic judge walkthrough and a public hosted instance.

What we learned

The best job-search automation is not automatic submission. It is reducing the number of decisions the user has to reconstruct while keeping high-stakes external actions explicit. Evidence provenance is also a user-experience feature: showing why a role fits and which résumé line supports a story makes the recommendation easier to trust and faster to use.

Codex was most valuable when product feedback crossed layers. A request such as “rank roles by fit” is not one UI sort; it changes persistence, pagination, cache identity, API semantics, edge cases, and tests. Keeping that context in one development task materially shortened the feedback-to-working-product loop.

What's next

  • More first-party company sources and user-selectable source packs.
  • Optional provider choices for public contact research while preserving the evidence and budget fences.
  • Outcome-calibrated recommendation weights once a user has enough data, without pretending small samples are causal.
  • A simpler guided Today flow that makes the next best action even more obvious on mobile.

Built with

Codex GPT-5.6 Python FastAPI PostgreSQL SQLAlchemy Alembic Next.js React TypeScript Docker Vercel Render Neon Google ADK Gemini Arize Phoenix SerpAPI

Built With

Share this project:

Updates