
AI/ML Engineer
Tiger Analytics · Jan 2024 – Present
- Architected a hierarchical multi-agent query resolution system with a Router/Orchestrator agent that classifies incoming queries as simple or complex, triggering lightweight or full multi-agent execution pipelines to optimize cost and response latency.
- Built RAG and web-search agents using pgvector for production vector storage and semantic retrieval (with ChromaDB/Qdrant supported for alternate deployments), with optimized embedding strategies to improve context relevance for orchestrator planning.
- Developed dynamic schema-filtering to inject only relevant tables/columns into the LLM context, reducing noise and improving SQL generation accuracy.
- Engineered real-time streaming for node-level response delivery and a Knowledge Base caching layer to cut redundant LLM calls, reducing user-perceived latency by 60% and lowering inference costs.
- Developed an automated evaluation framework for the SQL agent using an LLM-as-judge approach, ensuring consistent output quality at scale.
- Built LLM tool-calling infrastructure for automated scikit-learn model building and fine-tuning, enabling the agent to select the correct modeling tool and arguments and execute the full train/tune cycle autonomously.
- Designed and implemented a human-in-the-loop feedback system end-to-end, allowing users to steer agent execution mid-run; feedback dynamically revises the plan and is routed to the relevant downstream agent.
- Implemented strategies for managing intermediate artifacts (datasets, analysis outputs) outside LLM context to improve scalability and reduce token usage.
- Deployed the multi-agent platform on AWS using ECS Fargate and ECR for containerized, serverless orchestration of agent workloads.
- Evaluated DeepAgents as an alternative orchestration layer, benchmarking token consumption and inference cost against the LangChain/LangGraph stack and reverting based on quantitative analysis.
- Improved factual reliability of an existing RAG-based research chatbot by replacing hybrid BM25 + RRF retrieval with cross-encoder reranking, achieving the project's highest recorded citation accuracy and raising faithfulness by 27 percentage points (0.31 to 0.58) on a 90-query, 9-workflow evaluation benchmark.
- Engineered a bi-directional factual-contradiction guard to catch and fail LLM outputs that contradict retrieved source data, directly driving the faithfulness gains above.



