Fatim Majumder

About

ML/LLM infrastructure engineer who built and shipped a production internal evaluation + failure-analysis agent (Arthur AI / R3AI context) that orchestrated end-to-end benchmarks with deterministic lineage, regression detection, and root-cause reporting at 5,000+ benchmarks/week. Also built backend observability and data validation systems for analytics pipelines at FullStory processing ~3.4B weekly events, emphasizing schema validation, quarantine fallbacks, and idempotent operations.

Key Strengths

• Shipped production LLM evaluation/failure-analysis agent automating experiment-to-decision workflow
• Scaled system to support 5,000+ benchmarks per week with high reliability and graceful failure handling
• Built deterministic, reproducible run lineage (datasets/prompts/checkpoints/judges/configs/artifacts) improving accuracy from ~70% to 99.9%
• Reduced researcher debugging time by 40%+ and doubled throughput for validated model acceptance results
• Cut root-cause analysis cycle from ~2.1 days to ~2 hours via regression detection and run-diff tooling
• Identified a metrics inflation defect (reported as +6.1 pts; also cited +4.2 pts in another example) preventing faulty decision-making
• Designed robust evaluation/monitoring/feedback loops (offline slice evals + online anomaly detection + regression test encoding)
• Built data quality safeguards for messy, high-volume analytics pipelines (3.4B weekly events) using validation, quarantine, and idempotency

Experience

Education

Awards

• QuestBridge National Match Scholar
• Dean's List - Emory University
• Undergraduate Research Distinction
• HackDuke Finalist
• Selective Technical Leadership Fellow
• Dean's List

Skills

Languages