Training models, fighting bugs. Domain Expansion: Production AI

github

Shikhar Mishra  • Peerlist

Migrated from iterative LLM API call workflows to a tool-call-based agentic architecture, achieving 93% retrieval accuracy and pushing UAT ratings from 1.0 to 4.1 / 5.0 across enterprise evaluators. – Built multi-step reasoning and retrieval workflows for large technical document corpora (10,000+ documents) with streaming LLM responses and contextual memory management. – Implemented schema-aware retrieval tools, ACL-based search filtering, and citation-aware answer generation pipelines. – Participated in on-site deployment and infrastructure setup for local/offline enterprise deployment requirements.

Agentic RAG & Document Intelligence System

Designed backend infrastructure for document ingestion, replication-aware storage, and real-time filesystem synchronization across distributed node

Enterprise Document Intelligence Backend

Developed semantic enterprise search pipelines with personalization, question answering, query understanding, and Learning-to-Rank capabilities; NLP enhancements drove a 2× increase in active users and 120% longer average sessions.

Enterprise Semantic Search System

<p>▸ Built React 19 + FastAPI platform for electrical network analysis (14 – 70,000+ buses) with Cytoscape.js graph visualisation and MongoDB GridFS persistence. ▸ Developed LangChain ReAct agent with 60+ tools for natural-language querying across topology, power flow, contingency, OPF, and SCADA domains; RAGAS eval: ToolCallAccuracy 1.00 · Faithfulness 1.00. ▸ Integrated YOLO schematic-to-model pipeline; trained GNN (PyTorch) for fast state prediction; packaged 100+ MATPOWER benchmark cases (IEEE, RTS, ACTIVSg, PEGASE).</p>