Projects
4 repositoriesSIH RAG Assistant
Live demo availableProduction RAG chatbot answering queries over the Smart India Hackathon (SIH) knowledge base, combining semantic vector search with grounded LLM generation to reduce hallucination.
Hybrid retrieval in the API route: regex extraction resolves exact problem IDs (e.g., "SIH25050") directly from source JSON, while Vectra vector search with MiniLM-L6-v2 embeddings handles conceptual queries.
Zero-cost ingestion pipeline: parses raw JSON, generates on-device embeddings via Xenova Transformers with no external API calls, indexed into a persistent local vector store.
Tuned system prompts for Groq Llama 3.3 70B inference with real-time token streaming through Vercel AI SDK v6 for sub-second perceived latency.
A.R.G.U.S.
Demo video — Chrome Web Store listing pendingVoice-first AI browser automation agent. A LangGraph-orchestrated pipeline converts voice commands into typed automation actions, executed through a custom Playwright-style DOM runtime injected via content scripts to bypass Chrome extension sandbox limits.
Cut voice-command response time ~3x by re-engineering the Web Speech API capture layer with single-utterance mode, a 1.2s interim-transcript stability timer, and a 3s silence timeout.
Built a cascading element resolver (accessibility locators → role/text matching → AgentQL semantic resolution) plus a risk classifier and permission system that pauses purchases, deletions, and other high-risk actions for explicit user approval.
Edu-filter
Demo video — Chrome Web Store listing pendingChrome extension that reads YouTube video titles via the DOM, classifies each as educational or distracting via an LLM, and shows/hides videos accordingly in real time.
Multi-provider LLM support (OpenAI, Groq, Mistral) with user-supplied API key; key is stored locally client-side in the browser (not encrypted at rest).
HR Workflow Designer
Live demo availableGraph-based validation and simulation engine for organizational workflows — cycle detection, missing-link checks, and Start/End placement rules, running both client-side for instant feedback and server-side inside the simulation endpoint.
Mock REST API layer (GET /api/automations, POST /api/simulate) plus a simulation engine that walks the workflow graph from the Start node, returning step-by-step execution logs.
React Flow canvas with five typed, configurable node types (Start, Task, Approval, Automated Step, End) and JSON import/export for workflow persistence and replay.