AI Job-Title Normalization Pipeline

Description

Built a 4-step Agno workflow with 2 LLM agents that maps messy, free-text job categories (e.g. "Dev Front", "RRHH") to canonical O*NET occupation titles from a 923-title taxonomy, returning a corrected spreadsheet plus a full audit trail. A confidence-tiered routing system minimizes LLM cost: RapidFuzz auto-corrects exact/high-confidence matches (≥ 0.92), an LLM resolves semantic ambiguity (0.86–0.92), and low-confidence cases route to a human-in-the-loop review queue — the agent never overwrites data it can't verify against the source database. A multilingual translator agent (Spanish/Portuguese/French/Italian to English) runs as a fallback before escalation, recovering cases like "Cocinero" → "Cook" that plain matching would drop.

Highlights

• —4-step Agno workflow with 2 LLM agents mapping free-text job titles to a 923-title O*NET taxonomy, returning a corrected spreadsheet plus a full audit trail.
• —Confidence-tiered routing to minimize LLM cost: RapidFuzz auto-corrects matches ≥ 0.92, an LLM resolves ambiguity in the 0.86–0.92 range, and low-confidence cases go to a human-in-the-loop review queue.
• —Multilingual translator agent (ES/PT/FR/IT → EN) as a pre-escalation fallback, recovering cases like "Cocinero" → "Cook" that plain matching would drop.
• —Typed, auditable output enforced with Pydantic v2 schemas — every correction verified against the O*NET database before being written; zero free-text guesses.
• —Two-layer evaluation harness tracking precision (≥ 0.85) and hallucination rate (≤ 0.05), making agent quality objectively measurable and regression-safe.
• —Parallel mapper (ThreadPoolExecutor) for a ~3× wall-clock speedup, exposed via AgentOS (FastAPI) and a Chainlit UI with live step-by-step progress.

Architecture