my projects

deployed a multimodal voice AI agent combining voice and screen context using Whisper (STT), Gemini 2.0 Multimodal model, and gpt-4o-mini-tts (TTS) through EC2 on AWS with WebSocket streaming.

developed a perplexity sub-agent that can search the web and answer questions using the agent's context using perplexity MCP. orchestrated with langgraph workflows and DSPy for prompt optimization.

fine-tuned Qwen 2.5 (1.5B) model using Hugging Face TRL with LoRA-based SFT warmup & reinforcement learning via GRPO, aligning model outputs to first-principles reasoning across ~750 annotated interactions using RLAIF with curriculum scheduling.

built an agentic rag system that emulates charlie munger's thinking style (including his mental models) by integrating langgraph workflows, DSPy prompt optimization, and multi-agent orchestration (planner, retriever, mental model analyzer, synthesizer, verifier).

integrated DSPy for prompt optimization across planner, synthesizer, and verifier modules; retrieval pipeline uses FAISS with cross-encoder reranking, sustaining ~2.5s median end-to-end latency for multi-turn queries.

designed a RAG-based podcast summarization using OpenAI Whisper for transcription, Gemma3 for summarization, and hybrid RAG with semantic and BM25 retrieval through LangChain, with speaker diarization and DynamoDB storage.