I'm a Computer Science undergrad obsessed with LLM application engineering.

Shravani Patil  • Peerlist

Built FortifyFlow, a production-grade AI copilot that automates document processing, monitoring analytics, and reporting workflows for public health operations — eliminating 60%+ of manual knowledge work in food fortification programs. 🔧 Tech Stack: FastAPI · LangChain · OpenAI GPT-4 · ChromaDB · Pandas · Next.js 15 · Supabase · SciPy · Pydantic 📄 Document Intelligence — Built a multi-stage LangChain pipeline (PDF/DOCX → chunk → embed → extract) with Pydantic-enforced structured outputs for zero-failure extraction of actions, deadlines, risks, and stakeholders from policy documents. 📊 Monitoring Analytics — Engineered a CSV analytics service using Pandas + SciPy z-score anomaly detection (per-group statistical flagging) to identify critical district-level compliance outliers, paired with GPT-4 insight generation for plain-language root cause analysis. 🔍 Semantic Search (RAG) — Implemented a full RAG pipeline: OpenAI embeddings → ChromaDB vector store → cosine similarity retrieval → LLM synthesis, enabling semantic search across an entire document library. 📝 Reporting Assistant — Multi-source synthesis engine that combines document extractions + CSV insights into configurable stakeholder updates, donor reports, and field briefings with tone control. Architecture: Fully modular — separate frontend (Next.js 15), backend API (FastAPI), AI service layer (LangChain, ChromaDB), and data layer (PostgreSQL + vector store). Deployed on Vercel + Railway.

FortifyFlow — AI Workflow Copilot for Public 

• Gesture-Controlled Light System: Arduino + IR Sensor + OpenCV + Hand Gestures • Developed a system to turn lights on/off based on hand gestures • Used Arduino with IR sensors, LEDs, and OpenCV for gesture recognition • Enhanced understanding of embedded systems, computer vision, and human-computer interaction

Gesture controlled light system

Built an end-to-end EV fleet analytics product (Python, Pandas, GeoPandas, Streamlit) simulating 60-vehicle × 60-day operations across 10 urban zones; computed spatial accessibility and downtime risk scores using haversine distance + composite weighting, surfacing an 18% revenue loss from charger-access gaps