Kumar Saurabh

An AI-powered educational platform running on a decentralized local-network. Two laptops on the same Wi-Fi become a broadcast system: a Teacher Node captures speech, transcribes it locally, extracts concepts, and streams interactive HTML visualizations to students' browsers — no cloud, no app install required.

• Two-node LAN topology: Python/FastAPI teacher host streams to student thin clients over a browser URL with zero cloud deployment.
• Offline real-time transcription via faster-whisper running on host hardware — cost-free speech-to-text in background tasks.
• Concurrent LLM pipeline using LangGraph + asyncio.gather to parallelize fast text summaries and complex HTML visual rendering.
• Secure visual sandboxing: AI-generated HTML/CSS/JS injected into isolated iframes for safe dynamic animations on student interfaces.

A real-time multimodal surveillance assistant enabling natural voice interaction with live camera feeds — no traditional object-detection models required.

• Engineered a 3-tier Agent-to-Agent (A2A) microservices backend in Python/FastAPI: Vision Agent (frame analysis), NYC Data Agent (API integrations), and an Orchestrator (voice/text routing).
• Integrated NYC Open Data APIs to overlay live contextual city data (traffic feeds, incident reports) over real-time visual analysis.
• Built a low-latency WebRTC/Streaming frontend with React + Vite, using AudioWorklet and MediaRecorder APIs for full-duplex push-to-talk over WebSockets.

A schema-aware RAG chatbot enabling natural language querying over MySQL databases. Uses semantic search over DB schemas to generate accurate SQL, falls back to vector search or live web search when structured data is insufficient, and auto-adapts to schema changes — all powered by a local LLM with no cloud dependency.

• Query Classification — LLM-driven classifier (with regex fallback) routes each question to the right tool: SQL, vector search, or web search.
• Schema Layer — DB schema is introspected, embedded, and stored in a vector store; semantic search retrieves the most relevant tables before SQL generation.
• Text-to-SQL — Local LLM generates SQL from relevant tables with EXPLAIN-based validation and one-retry self-correction on errors.
• Row-Level Vector Store — Text columns from MySQL rows are chunked and embedded for semantic similarity search when SQL returns no results.
• Web Search Fallback — Provider-agnostic layer (Perplexity → Tavily → DuckDuckGo) handles questions outside the database.
• CDC Pipeline — Change Data Capture (polling) keeps the vector store in sync with live DB changes automatically.
• Schema Watcher — Detects ALTER / CREATE / DROP TABLE events and re-indexes only the affected tables.
• Chat API + UI — FastAPI server with SQLite-backed conversation history and a browser-based chat interface.

Led end-to-end development of a comprehensive bond trading ecosystem designed for institutional-grade price discovery and portfolio optimization across Govt, Corporate, Municipal, and REIT/InVit instruments.

• Architected a hybrid dual-database system: MySQL for structured transactional integrity and MongoDB for flexible, high-volume security metadata.
• Scaled automated ingestion engines using APIs, web crawlers, and LLMs to aggregate real-time data across multiple fixed-income instrument types.
• Directed implementation of Quantitative and ML models for risk-based pricing and default prediction, powering a proprietary internal scoring system.
• Built portfolio construction tools for Principal Protection, Target Returns, and Liquidity Management mandates.
• Spearheaded UI/UX strategy with React and React Native for seamless cross-platform web and mobile delivery.