Nishchal P R.AI Engineer · GraphRAG & Agents

Vector search collapses on multi-hop financial questions where answers live in entity relationships, not single chunks.

• →LLM-driven entity-relation extraction for dynamic knowledge graph construction
• →Hybrid retrieval combining semantic similarity with graph traversal
• →Graph reduction pipeline compressing ~1M nodes → ~10K high-signal nodes
• →Mitigated entity duplication, noisy relations, signal loss during pruning

• Multi-hop reasoning on questions vector RAG couldn't answer
• ~100× compression with preserved retrieval precision
• Latency-stable under iterative graph updates

Support teams burn hours stitching together shipment events, logs, and tickets to root-cause a single delayed delivery.

• →Agentic tool-calling with hybrid retrieval over shipment events, logs, and support records
• →ReAct (Reason + Act) orchestration loop for autonomous anomaly diagnosis and RCA
• →Evaluation pipeline benchmarking tool-selection accuracy, reasoning trajectory, and latency

• Autonomous root-cause analysis from raw operational signals
• Tool-selection accuracy measured and tracked
• Production-grade latency and reliability targets

Static single-model deployments waste compute budget on simple queries and underperform on complex ones.

• →Heuristic complexity estimator scoring incoming queries across multiple dimensions
• →Dynamic routing across LLM tiers (small/medium/large) based on complexity score
• →Cost-quality tradeoff optimization with configurable quality floors
• →Latency and cost telemetry per routing decision

• Significant cost reduction without measurable quality degradation
• Adaptive routing that improves with query volume
• Full observability on cost-per-query

Naive vector RAG fails on complex queries requiring context synthesis across multiple documents — and there's no reliable way to measure how bad it is.

• →Hybrid retrieval: dense + sparse (BM25) with contextual reranking
• →Evaluation pipeline: RAGAS metrics (faithfulness, answer relevancy, context precision/recall)
• →Systematic benchmarking of retrieval quality and response reliability
• →Chunk strategy optimization and embedding model comparison

• Measurable improvements in faithfulness and context precision
• Production-grade evaluation harness reusable across projects
• Systematic approach to retrieval quality