Amit kumar singh

Architected an autonomous, hallucination-free Agentic RAG system for Indian jurisprudence enforcing strict legal hierarchy (Constitution Supreme Court High Court) via a 13-step deterministic Python state machine no LangChain, no framework lock-in. Engineered a multi-tier hybrid search pipeline (Dense Vectors, Sparse BM25 with morphological stemming, Reciprocal Rank Fusion, and Cross-Encoder reranking) to achieve ultra-high precision on complex legal texts. Pushed jurisdiction authority logic into Vector DB metadata (SC weight=1.0, HC=0.7, District=0.2), making legal precedence algorithmic rather than LLM prompt-dependent. Implemented a kill-switch abstention engine with domain-specific rules detecting repealed statutes (IPCBNS), wrong jurisdictions, and non-binding rulings ensuring zero hallucination liability in production. Optimized to run fully locally on RTX 5060 Ti (16GB VRAM) using vLLM with AWQ 4-bit quantization; zero data sent to external APIs.

Built a local, zero-data-leakage AI inference platform for regulated industries (Finance, Insurance, Healthcare) with resource-aware infrastructure that auto-detects hardware and falls back between GPU (HuggingFace) and CPU (llama.cpp) backends. Engineered a document-scoped inference gateway with dynamic LoRA/QLoRA/DoRA adapter loading at runtime, preventing generic model hallucinations on proprietary internal policy documents. Designed multi-teacher ensemble generation strategies (Vote / Concat / Rerank) to improve deterministic accuracy for audit-heavy compliance workflows. Built a human-in-the-loop curation pipeline via Label Studio webhooks with full data provenance, versioned taxonomies, and audit logs. Deployed on a distributed microservices stack: FastAPI + Celery + Redis + PostgreSQL + MinIO, containerized via Docker/Podman with single-command make targets for GPU and CPU variants.