Amogh

Multi-Agent Clinical AI System

Architected a multi-agent clinical AI system that synthesizes messy longitudinal health records (PDF/JSON) into structured EHR-ready discharge summaries. - Enforced zero-fabrication guardrails with a source-citation verification module; the agent triggers escalate() to alert clinicians of contradictions and unmanaged events instead of hallucinating. - Designed an In-Context Correction Memory Loop with a simulated critic to inject localized corrections upstream and prevent recurrent token-matching errors. - Optimized safety metrics to 1.00 safety recall on held-out patient sets, trapping critical allergen contraindications, missing rationales, and truncations. Tech: Python, LangGraph, LangChain, Pydantic, FastAPI, MongoDB

View details

Agentic AI RAG System with Caching

Hybrid agentic RAG pipeline combining BM25 lexical retrieval, dense vector search (all-MiniLM-L6-v2), and cross-encoder re-ranking to select the most relevant academic context with high precision. - LangGraph-based execution graph with conditional routing, retry logic, and state propagation for deterministic orchestration of retrieval, generation, and evaluation. - End-to-end agentic JSON generation and validation loop: LLM generates structured outputs, a Pydantic validator enforces schema correctness, feeds errors back, and ensures production-ready responses. Tech: Python, LangChain, LangGraph, BM25, ChromaDB, Pydantic, Cross-Encoder

View details

Startup Intelligence & Funding Analytics Pipeline

Built an end-to-end startup intelligence pipeline that ingests articles from multiple RSS sources including TechCrunch, Inc42, The Hindu, and YourStory to identify funding, acquisition, and merger events. - Developed an LLM-powered extraction framework to classify article relevance and transform unstructured news content into structured datasets containing startup, funding round, investor, and acquisition details. - Engineered a multi-stage enrichment workflow integrating LinkedIn, CoreSignal, Explorium, and Tavily to gather founder profiles, company intelligence, investor information, and historical funding data. - Implemented stealth browser automation using Puppeteer and Selenium for authenticated data collection, dynamic content extraction, and company website discovery across gated and JavaScript-heavy sources. - Created a scalable async processing pipeline that generated analysis-ready datasets with startup metadata, funding amounts, investor details, and founder insights for market intelligence and research applications. Tech: Python, FastAPI, Azure OpenAI, Selenium, Puppeteer, RSS, LinkedIn API

View details