Project 01 / 06
Full-stack tool that compares résumé and job-description skills, surfaces matched and missing skills, and tracks recurring gaps across saved roles.
One captured analysis result
The interface separates matched skills from skills listed in the job description but missing from the résumé. A warning reminds users to review the result.
The screenshot does not show the full inputs, the analysis method used or whether the result was saved. The skill classifications still need review.
A rule-based path creates the comparison independently of optional AI availability.
The matching code lowercases text, maps skill aliases to a common vocabulary and uses boundary-aware regular expressions to find skills. It removes duplicates, then compares the résumé and job-description skill sets.
The skill vocabulary contains 23 categories, 250 canonical skills and 99 alias groups. These are vocabulary counts, not measures of matching accuracy.
| Stage | How it works | Limitations |
|---|---|---|
| Normalize | Lowercase text and map documented aliases to canonical skill names | No general language understanding is implied |
| Detect | Use boundary-aware regular expressions to avoid simple substring matches | Not semantic matching, embeddings, or vector search |
| Compare | Intersect résumé and job skill sets for matches; retain job-only skills as missing | No candidate ranking or fit formula |
| Configure | 23 categories · 250 canonical skills · 99 alias groups | Configuration counts, not performance metrics |
The skill-detection fixtures used 98 alias groups. The vocabulary described above contains 99.
The OpenAI path adds structured analysis without becoming the workflow's only route to a result.
When configured, the FastAPI service calls the OpenAI Responses API with a strict structured-output schema for matched skills, missing skills, and a summary. The request disables provider-side response storage.
The web layer canonicalizes and deduplicates returned skills, reconciles them with deterministic overlap, and prevents a skill from remaining in both matched and missing sets. If AI is disabled, fails, or returns an unusable payload, the route returns the rule-based result instead.
Public product framing
| Concern | Implementation |
|---|---|
| Output shape | Strict structured response for skills and summary |
| Vocabulary | Canonicalization and deduplication |
| Set consistency | Deterministic overlap reconciles matched and missing skills |
| Service failure | Automatic fallback to rule-based analysis |
Structured responses, vocabulary checks and fallback handling do not ensure correct matches. AI output is planning guidance and needs human review.
The hosted workflow separates authenticated request handling, analysis, document extraction, optional AI, and explicit persistence.
Authenticated Next.js routes proxy analysis requests to FastAPI. The backend exposes deterministic analysis, document extraction, and optional AI endpoints; PDF, DOCX, TXT, and Markdown content can be converted to transient text, while OCR is not implemented.
The request handlers include size limits and timeouts.
Analysis workspace — input surfaces
| Stage | How it works |
|---|---|
| Browser → Next.js | Authenticated server routes receive and validate analysis requests |
| Next.js → FastAPI | Server-side proxy calls deterministic, extraction, or optional AI services |
| Document extraction | PDF, DOCX, TXT, and Markdown text extraction |
| Analysis → save | Persistence is a separate explicit action using structured fields |
An explicit save turns one transient result into structured records that can be revisited across roles.
Saving writes structured analysis metadata and matched and missing skills to Supabase. Row-level policies restrict each user’s access to their own records. This save path excludes raw résumé and job-description text.
The app supports searching, filtering, comparing, exporting and deleting saved analyses. Recurring-gap percentages count the unique saved analyses that include each skill, so they describe one user’s saved roles rather than the wider labor market.
| Feature | How it works |
|---|---|
| Structured save | Run, role, matched-skill, and missing-skill records |
| Owner scope | Supabase row-level policies tied to the authenticated user |
| Saved comparison | Search, filter, compare, export, and delete controls |
| Recurring gaps | Unique saved-analysis counts grouped by canonical skill |
The matching logic was tested with fictional job-role examples, negative controls, and skill-alias checks. The project also went through QA and launch checks.
| Test | Recorded result |
|---|---|
| Skill-detection fixtures | 21 fictional role cases · 19 negative controls · 49 alias-driven detections · 0 final fixture failures |
| QA checklist | 17 of 17 recorded checks passed |
| Launch checks | 7 of 7 recorded checks passed, plus manual review |
These checks took place at different stages of development. They do not measure real-world matching accuracy, hiring outcomes or reliability at scale.
JobFit started as a rule-based command-line tool and grew into a web app with saved results, profiles, privacy controls, and optional AI analysis. I worked on its development with substantial assistance from AI coding tools.
