Why a guard after the prompt
A prompt that says 'do not invent facts' is a request. The fact guard is a check: it compares the rewrite against the original resume and throws the rewrite away if anything new appears.
Jobs with Joe
A job board for Indian freshers: twelve source adapters, explainable rule-based classification, and AI resume tools behind a fact guard that rejects anything the candidate did not write.
- Role
- Designed and built
- Where
- Personal project
- When
- 2026
- Access
- Private repo · live product
Problem
Freshers in India wade through listings that are mislabelled, duplicated, or scams, then rewrite the same resume for every application. AI can help with the writing, but a resume that gains a skill or a number the candidate never had is worse than no help.
01
Twelve sources, one clean plan
Adapters collect jobs from applicant-tracking APIs and structured job markup, respecting robots.txt and sanitizing every field. A pure, I/O-free plan normalizes, validates, classifies, and deduplicates, then applies everything in one transaction.
02
Classification you can explain
An ordered rule set labels each job's experience level and returns the IDs of the rules that fired. 'Unknown' is a normal answer and is never shown as fresher-friendly.
03
AI is optional
A model router uses Gemini Flash, retries on a lighter model after a rate limit or timeout, and can use OpenAI instead. With no key at all, every AI feature falls back to a deterministic version.
04
A fact guard after generation
Prompts forbid new employers, numbers, or skills and turn missing facts into placeholders. Then a post-check rejects any rewrite that adds a number, a known skill, or a proper noun that is not in the original resume.
05
The candidate approves every line
Rewrites are accepted line by line. Applying one checks the source line has not changed and escapes LaTeX, so a rewrite cannot inject commands into the generated PDF.
06
Matching with minimal data
Optional embeddings power job matching, built from skills and preferences only: no name, grades, or pay go to the embedding model.
07
A scam shield with no model in it
Lookalike domains, punycode, fee or UPI requests, and personal email addresses are caught by deterministic rules, logged anonymously, and rate limited.
Measured
- Job source adapters
- 12
- Unit and end-to-end tests
- 1,305
- Planned pgTAP database assertions
- 826
- CI workflows
- 11
- Database migrations with row-level security
- 37
How it's verified
- Harness
- One OpenAI-compatible client with a model fallback chain, a health probe, and a deterministic fallback for every AI feature; prompts and outputs are never logged.
- Guardrails
- Fact rules in every prompt plus a post-generation guard that rejects new numbers, skills, or names; schema-validated JSON output; rate-limited extension tokens with an origin allow-list; data minimisation for embeddings; line-by-line user acceptance.
- Testing
- About 1,305 tests (574 API, 581 web unit, 68 Playwright, 82 extension) plus 826 planned pgTAP assertions, run by separate CI workflows for web, API, database, extension, and LaTeX.
In depth
A browser extension that cannot be borrowed
The Apply Copilot extension authenticates with its own bearer token, is allowed only from extension origins, and each user has a quota enforced in the database.
Tests at every layer
Pure functions get unit tests, the web app gets Playwright flows, the database gets pgTAP assertions run against a throwaway local Supabase in CI, and the LaTeX service gets its own build check.
Decisions
- 01AI as an optional feature layer with a deterministic fallback for every feature.
- 02Rule-based, explainable classification instead of an LLM.
- 03Row-level security first, so reads work with the public key and nothing else.
Stack
Next.js · TypeScript · FastAPI · Supabase Postgres (RLS) · pgTAP · Gemini / OpenAI · Chrome extension (MV3) · Playwright · Vitest · Docker LaTeX