Vismay Devjee

About

Asset Management Risk professional at Fidelity Investments who built and productionized an agentic RAG platform enabling compliance and analysts to query 10,000+ fund documents with cited answers in seconds. Implemented structure-aware semantic chunking (AWS Textract), hierarchical retrieval, and hybrid search to raise accuracy from 68% to 94%, and built an evaluation framework tracking accuracy/latency/cost/hallucinations—delivering 40+ hours/month saved and zero critical production failures.

Experience

Education

Key Strengths

• Built and deployed an end-to-end agentic RAG system for querying 10,000+ financial fund documents
• Improved retrieval accuracy from 68% to 94% via hybrid retrieval and iterative tuning
• Designed structure-aware semantic chunking for dense financial tables using AWS Textract and custom Python
• Implemented hierarchical retrieval (category → document → granular chunks) with rich metadata for context preservation
• Established guardrails and out-of-scope query handling to reduce hallucinations in production
• Created an LLM evaluation system tracking accuracy, latency, cost per query, and hallucination rate (citation verification)
• Strong reliability engineering: schemas, confidence thresholds, fallbacks, human-in-the-loop, unit + integration testing
• Cross-functional delivery: ran weekly working sessions with non-technical stakeholders and coached users on prompting
• Demonstrated measurable business impact (94% accuracy, 40+ hours saved monthly, zero critical production failures)

Reference Highlights

• Proactive iteration based on feedback
• Comfortable presenting to stakeholders at all levels
• Strong communication with both technical and non-technical audiences
• Able to translate technical concepts into business context
• Responsive in resolving what worked/didn't work to meet requirements
• Instrumental in creating an LLM-based evaluator for an operational risk concern
• Respected by team members
• Pleasure to work with
• Clear and concise communicator in most areas
• Able to explain difficult concepts (agentic AI) to non-technical audiences reasonably well
• Works effectively with guidance from senior engineers and architects
• Effective collaboration with business partners during beta testing