Selected Work
Systems I've built
Veritas-Agent
An 8-node LangGraph pipeline (classification → retrieval → reasoning → validation → retry) with dynamic routing, hybrid retrieval (BM25 + FAISS + RRF + cross-encoder reranking), and an adversarial validation agent run on a separate prompt from the reasoning agent to avoid the confirmation bias of a model grading its own answer. Confidence scores route each query to validation, retry with a reformulated query, or a targeted single-question clarification, and short/long-term memory (Redis + FAISS) carries context across turns and sessions.
81% task completion and 0.91 faithfulness on a 200-query benchmark; validation alone accounts for a 13-point faithfulness gain, and failure-aware retry lifted completion from 67% to 81% while cutting P99 cost from ₹4.20 to ₹1.10.
PythonLangGraphOpenAI GPT-4oFAISSRedisLangSmithRAGASFastAPIStreamlitGrounded-SQL
Fine-tuned Mistral-7B-Instruct with QLoRA (4-bit NF4, rank ablated across 8/16/64) on 8,659 Spider question-SQL pairs (166 schemas, 138 domains, split 8,080 train / 495 validation / 1,034 test), evaluated by executing generated SQL against the live database and diffing result rows rather than comparing strings, then isolated the generalization bottleneck via a four-way ablation (base / RAG / fine-tuned / fine-tuned+RAG).
Execution accuracy improved 53.3% → 70.2% (+32% relative) on 1,034 unseen queries across 166 databases, with hybrid schema linking (BM25 + FAISS + RRF) eliminating hallucinated-column errors entirely.
PythonMistral-7BQLoRAFAISSFastAPIRedisDockerGradio