Samarth Singh

I’m an AI/ML engineer focused on turning messy, real-world documents into structured, trustworthy outputs using LLM pipelines. I enjoy building systems where performance and correctness are measurable, not just promised.

In my current role, I architected a vision extraction pipeline for multi-page architectural PDFs using GPT-5, cutting latency from ~7 minutes to ~90 seconds with concurrent batch dispatch, structured output enforcement, and retry handling. I also built an intake pipeline that combines non-AI heuristic extraction (regex/Docling) with conditional GPT-5 fallback, achieving confidence > 0.85 and near-complete field extraction while only invoking vision when needed.

On the product side, I built DoCopilot, a full-stack RAG document Q&A system using FastAPI + Next.js with Qdrant hybrid search (BM25 + dense vectors), RRF fusion, and cross-encoder reranking. I hardened it with prompt-injection detection, PII redaction, and source-grounding checks, and validated retrieval quality with an LLM-as-Judge study (89.2% correctness, 90.5% relevance, 100% source rate, 2.86s avg latency).

I’m also deep in agentic workflows—Argus is a LangGraph supervisor multi-agent engine that autonomously synthesizes cited reports, reducing manual research time from hours to 30–90 seconds with checkpointing and LangSmith observability. For GenAI research, I fine-tuned FLAN-T5-base with LoRA, built a reproducible evaluation suite (ROUGE/BERTScore/METEOR/ BLEU), and deployed an interactive Gradio summarizer on Hugging Face.