Amit Gangane

About

LLM/agentic systems builder who shipped production workflows at Angel Flight West and Eureka AI, combining LangGraph + RAG (Postgres/pgvector) with strong observability (LangSmith/Langfuse). Delivered large operational gains (address lookup cut from 10 minutes to 60 seconds; accuracy to 92%) and has a track record of quickly stabilizing customer-critical pipelines (Pydantic-enforced JSON for ETL) while partnering with sales/ops to drive adoption.

Experience

Education

Key Strengths

• Took an LLM agent from prototype to production with measurable impact (10 min to 60 sec lookup; 30% to 92% accuracy)
• Systematic LLM debugging using observability/tracing (LangSmith/Langfuse) to pinpoint failures
• Improved extraction precision by 23% via prompt iteration and few-shot examples for messy real-world data
• Improved retrieval performance by tuning embeddings/chunking (reported 95% retrieval accuracy)
• Rapid incident response: fixed malformed JSON breaking ETL by enforcing Pydantic structured outputs (deployed within hours)
• Strong technical communication: practical developer-focused demos with code, architecture, and design tradeoffs
• Cross-functional partnership with sales/ops to align technical outputs to customer needs and drive adoption
• Built and deployed production agentic LLM workflow classifying 125K+ unstructured network traffic logs; reduced manual labeling work by 90%+
• Improved unknown-domain classification accuracy via RAG + dynamic search grounding (A/B test: 60% -> 89% on 1,000 logs)
• Enforced reliable structured outputs using Pydantic JSON schemas to eliminate formatting errors in downstream ETL
• Designed scalable cloud data architecture (BigQuery streaming + Cloud Functions) processing telco data across US/HK/UK
• Used LangSmith tracing/evaluation to identify failure modes and improve extraction accuracy by 23%
• Established measurable success metrics and built ground-truth labeled evaluation sets (e.g., manually verified 200 facilities) to track precision/recall
• Delivered AI solution with non-technical stakeholders by shadowing users and prioritizing transparency/verification; reduced lookup time from 10 minutes to 60 seconds with ~90% accuracy