Inspiration
Applying to several jobs should not require repeatedly reading the same CV, copying job descriptions, guessing ATS keywords, rebuilding documents and re-entering the same information on every career portal. Existing tools usually automate one fragment and hide the decisions they make. I built JobAuto Studio to make the complete process visible and candidate-controlled.
What it does
JobAuto turns a trusted candidate profile into an observable application workflow:
- Import a LaTeX CV or create a structured profile.
- Review extracted experience, skills, projects and protected facts.
- Choose what Codex may preserve, reframe or adapt.
- Search current job pages in batches.
- Deduplicate, filter and rank opportunities against the candidate's preferences.
- Analyze the full offer for role fit, ATS requirements and supporting evidence.
- Generate a tailored CV and cover letter.
- Compile and inspect the real one-page PDFs.
- Run an independent final review and repair only the document that needs it.
- Create a hash-verified packet for the user's authenticated Chrome session and store the resulting receipt.
The user can inspect the original and tailored documents, the ATS coverage, the selected evidence, agent events, warnings, file hashes and application status from one local interface.
How we built it
JobAuto Studio is a local Python and FastAPI application. Pydantic models define the contracts shared by discovery, strategy, document generation, review, rendering and submission. The filesystem stores immutable profiles, full job descriptions, generated artifacts and receipts, while an Excel tracker remains available for campaign reporting.
Codex is the orchestration and browser-execution layer. GPT-5.6 reads complete offers, identifies the actual role and its important requirements, maps them to candidate evidence, selects the most relevant angle and projects, writes the CV and letter, and performs an independent rendered-document review. Codex web search discovers current openings, while the Codex Chrome Extension executes approved packets in the user's authenticated browser.
The system deliberately separates two kinds of work:
- Agentic decisions: understanding the role, choosing evidence, adapting the narrative and evaluating relevance.
- Deterministic guarantees: preserving the LaTeX structure, compiling PDFs, enforcing one-page limits, checking protected facts, calculating hashes and recording receipts.
This separation keeps the result adaptable without giving the model control over document integrity.
Challenges
The hardest problem was not generating text. It was preserving a candidate's real CV layout while allowing meaningful adaptation, then proving that the generated files were the exact files handed to Chrome. We also had to distinguish unsupported evidence from ordinary ATS gaps so that secondary missing keywords did not block otherwise strong applications.
Browser automation introduced another boundary: login state and employer portals belong to the user's Chrome session, not the FastAPI server. JobAuto therefore produces a machine-readable handoff, verifies the files again immediately before upload, and records confirmation separately from document generation.
Accomplishments
The demonstrated synthetic campaign evaluated 9 offers, selected 5 applications, generated 10 one-page PDFs, ran 5 independent final reviews and persisted 5 verified sandbox receipts. The demo also shows source-versus-tailored documents, ATS evidence, agent traces and exact artifact hashes.
Most importantly, JobAuto is not a black-box resume writer. It is a reusable workflow in which candidates define the facts and adaptation freedom, Codex performs the contextual work, and every important transition remains inspectable.
What we learned
High-quality automation needs both autonomy and boundaries. Rigid keyword templates produce generic documents, while unconstrained generation risks changing facts or layout. The useful middle ground is a typed candidate context, explicit adaptation permissions, full-offer grounding, independent review and deterministic artifact checks.
Codex was also valuable beyond implementation: it built and tested the application, operated the end-to-end workflow, inspected the rendered PDFs, researched offers and controlled the authenticated Chrome handoff.
What's next
The next priorities are reducing latency through shared context and conditional review, expanding portal compatibility, improving the visual block editor, and packaging the local installation for more Codex users.
Log in or sign up for Devpost to join the conversation.