Chinmaye H.G.

Built with Next.js 16, TypeScript, Express 5, and PostgreSQL (pgvector) via Drizzle ORM. Utilizes LangGraph, Groq, and Jina AI embeddings for end-to-end RAG pipelines and automated human-handoff workflows.

Standard customer support tools lack contextual source-cited intelligence. Architecting a multi-tenant RAG system with optimized 32-chunk batch processing cuts embedding API calls by 97% while ensuring secure, isolated customer data handling.

Built with Next.js 16, TypeScript, FastAPI, and MongoDB, with Clerk for auth and LiveKit Cloud for real-time WebRTC audio. A live voice agent runs a Groq Whisper, Groq LLM, and Deepgram TTS pipeline, saves the transcript to MongoDB turn by turn, and auto-scores each candidate with schema-validated LLM output.

Real-time voice is too heavy for a 512 MB free-tier host, so I split the system into a lightweight FastAPI API and a separate voice agent on LiveKit Cloud, with all speech and LLM work offloaded to Groq and Deepgram. Atomic MongoDB status transitions stop an invite being started twice, per-company data isolation keeps recruiters from seeing each other's candidates, and a release-on-failure path means an infrastructure error never uses up a candidate's one interview.

Powered by FastAPI, LangGraph, React 19, and PostgreSQL (pgvector). Implements a 5-stage ingestion pipeline with LLM-driven contextual metadata enrichment and dynamic retrieval verification loops.

Unverified RAG responses frequently suffer from silent hallucinations and high latency. Introducing an automated verification node and an intent router short-circuits unnecessary vector searches, reducing hallucination rates by 85% and cutting token consumption by 60%.

Executes parallel requests across Gemini, LLaMA, DeepSeek, Qwen, and Gemma model providers. Aggregates stream responses into a unified side-by-side verification interface.

Single-model LLM workflows suffer from silent hallucinations. Querying models in parallel lets users compare consensus and confidence across distinct architectures before acting on generated output.

Integrates Clipdrop AI API endpoints into an authenticated user workflow with persistent transformation history and asset management.

Offloading complex computer vision model execution to specialized serverless AI endpoints ensures ultra-low latency client interactions and scalable image transformation pipelines.

Parses long-form technical video streams into modular chapters, syllabus hierarchies, and interactive study modules.

Raw video feeds lead to unstructured learning. Extracting structured metadata into defined learning paths drastically improves content retention and scannability.