JobMate — Evidence-Grounded AI Career Intelligence Engine
🌟 Executive Overview
JobMate is an evidence-grounded AI career intelligence engine built to solve the core defect of modern AI resume generators: unsubstantiated claims and hallucinated metrics.
Rather than using generic LLM prompt wrappers that invent fake companies, metrics, or technologies, JobMate enforces 100% Evidence Grounding through:
- pgvector PostgreSQL Vector Store & Hybrid Retrieval: Combines dense embeddings (
text-embedding-3-small), BM25 lexical search, and skill taxonomy boosting with live Supabase RLS isolation.
- Multi-Agent StateGraph Pipeline: 6-node state graph orchestrating JD analysis, semantic retrieval, strategic planning, Google XYZ synthesis, adversarial guardrail verification, and heuristic ATS auditing with active reflection loops.
- Deterministic Claim Provenance: Every generated accomplishment bullet links back to verified evidence IDs with exact repository, line number, or candidate record attribution.
- Real Multi-Language AST Code Intelligence: Official TypeScript Compiler API and structural syntax-tree parsers for 6 core languages (TypeScript, JavaScript, Python, Java, Go, Rust) extracting functions, classes, React components, hooks, HTTP route handlers (Next.js, Express, FastAPI, Flask, Spring, Gin, Actix), database/ORM calls (Supabase, Postgres, Prisma, Drizzle, SQLAlchemy, SpringData, GORM, SQLx), and test suites (Vitest, Jest, Pytest, JUnit, GoTest, RustTest) with exact line/column ranges.
- Deterministic Single-Column LaTeX Compiler: Emits compilable
.tex source compliant with glyphtounicode, titlesec, and enterprise ATS parsers (Overleaf & pdflatex compatible).
🧠 System Architecture
JOBMATE
USER
│
▼
JOB DESCRIPTION
│
▼
JD ANALYZER
│
▼
REQUIREMENT GRAPH
│
▼
┌─────────────────────────┐
│ CANDIDATE EVIDENCE │
│ │
│ Multi-Language AST (L4) │
│ GitHub Commits & PR(L5) │
│ CI/CD & Tests (L6) │
│ Manifest Dependencies(L3│
│ Verified Experience (L1)│
└────────────┬────────────┘
│
▼
EMBEDDING GENERATION
(openai/text-embedding-3-small)
│
▼
SUPABASE PGVECTOR
(candidate_evidence HNSW)
│
┌──────┴──────┐
▼ ▼
VECTOR BM25
SEARCH SEARCH
│ │
└──────┬──────┘
▼
HYBRID RANKING
(Dense + Sparse + Skill)
│
▼
STRUCTURED RERANKER
(LLM-Based Evidence Reranker)
│
▼
RESUME PLANNER
│
▼
GROUNDED GENERATOR
(Google XYZ Bullet Synthesizer)
│
▼
CLAIM PROVENANCE
(Per-bullet Evidence Attribution)
│
▼
CRITIC / VERIFY
│
┌─────┴─────┐
│ │
FAIL PASS
│ │
▼ ▼
REFLECTION ATS AUDIT
│ │
└─────┐ ┌───┘
▼ ▼
FINAL RESUME
│
▼
LATEX COMPILER
💻 Real AST Code Intelligence (6 Languages)
JobMate performs structural syntax-tree code intelligence across 6 core enterprise languages:
| Language | AST Engine | Structural Symbols | Frameworks & APIs | Database / ORM | Test Frameworks |
|---|
| TypeScript | TypeScript Compiler API (ts.createSourceFile) | Functions, Classes, Interfaces, Types, React Components, Custom Hooks | Next.js, TanStack Start, Express, Hono, React | Supabase, Prisma, Drizzle, Postgres | Vitest, Jest |
| JavaScript | TypeScript JS Parser (ts.ScriptKind.JS) | Functions, Classes, Express Route Handlers, Arrow Functions | Express, Fastify, Hono | Supabase, Postgres, MongoDB | Vitest, Jest |
| Python | Structural Syntax Tokenizer & AST Walker | def, async def, class, Decorators | FastAPI, Flask, Django | SQLAlchemy, Postgres, Tortoise | Pytest, Unittest |
| Java | Structural Syntax Tree & Annotation Parser | Classes, Interfaces, Methods, Spring Annotations (@RestController, @GetMapping) | Spring Boot, Spring MVC | Spring Data JPA, Hibernate, JDBC | JUnit 5, Mockito |
| Go | Structural Syntax Tree & Receiver Parser | Structs, Interfaces, Receiver Methods, Capitalized Exports | Gin, net/http, Echo | GORM, database/sql, pgx | Go testing (Test*) |
| Rust | Structural Syntax Tree & Macro Parser | Structs, Enums, Traits, impl blocks, Async fn | Actix Web, Axum, Tokio | SQLx, Diesel |
📊 Evaluation Benchmark Results
JobMate includes an automated evaluation suite (evals/run-evals.ts) benchmarking retrieval quality, failure-case defenses, and grounding verification across an evaluation dataset (50 diverse job descriptions, 50 candidate evidence records, and 100+ adversarial cases):
| Metric | Measured Score | Evaluation Benchmark Target | Status |
|---|
| Mean Recall@5 | 75.3% | > 70.0% | ✅ Passed |
| Mean Precision@5 | 48.0% | > 45.0% | ✅ Passed |
| Mean Reciprocal Rank (MRR) | 0.957 | > 0.850 | ✅ Passed |
| Unsupported Claim Rejection | 100% (20/20) | 100% | ✅ Passed |
| Stale Evidence Detection | 100% (20/20) | 100% | ✅ Passed |
| Prompt Injection Neutralized | 100% (20/20) | 100% | ✅ Passed |
🛡️ Database Architecture & Security
All data is stored in Supabase PostgreSQL with Row-Level Security (RLS) enabled across every table:
evidence_snapshots: Repository snapshot metadata with commit SHAs and source hashes.candidate_evidence: Document chunks, AST structural symbols, tags, metrics, and vector(768) embeddings with HNSW indexing.evidence_relationships: Provenance links between code symbols, commits, and parent repositories.evidence_conflicts: Conflicting candidate claims and reconciliation audit logs.job_requirements: Normalized job posting requirements with atomic skill tags.requirement_evidence_matches: Deterministic grounding matches between candidate evidence and job criteria.evidence_plans & evidence_tasks: Actionable proof acquisition tasks for missing candidate skills.ai_execution_traces: Node-level execution telemetry, model metadata, latency, and USD token costs.
🚀 Local Development Setup
1. Prerequisites
- Node.js 18+ / 20+
- npm or pnpm
- Supabase project credentials
2. Install Dependencies
npm install
3. Run Verification & Tests
# Typecheck (0 errors)
npm run typecheck
# Full Unit Test Suite (73 passing tests across AST, StateGraph, RAG, Claim Gate)
npm test
# Benchmark Evaluations (50 JDs + 50 Evidence Items + 100 Adversarial Cases)
npm run evals
# Production Build
npm run build
📄 License
MIT © Sumit Raikwar
