Technical Arsenal

The tools and technologies I use to build intelligent systems.

Agentic Frameworks

  • Orchestration: I use LangGraph & LangChain to build cyclic graph-based flows and manage state.
  • Models: I leverage Gemini 2.0 Flash (Function calling), Anthropic Claude 3.5 Sonnet (Coding), and Llama 3 (for local reasoning).
  • Tooling: I build custom tools using SerpAPI (Web), Gmail API (Auth), and file system agents.

Advanced RAG

  • Vector Stores: I scale production systems with Pinecone (Serverless) and Weaviate (Hybrid search indices).
  • Retrieval: I implement Reciprocal Rank Fusion (RRF), ColbertV2 reranking, and Parent-Child chunking strategies.
  • Embeddings: I fine-tune SentenceTransformers to maximize dimension efficiency and relevance.

Engineering Core

  • Backend: I deploy Async FastAPI/Flask microservices, use Pydantic for validation, and WebSockets for streaming.
  • ML Ops: I use Docker containerization for reproducible environments and managing inference endpoints.
  • Deep Learning: I implement PyTorch reinforcement learning agents (DQN, PPO) in my custom Gymnasium environments.

Niche Integrations

  • Voice Pipeline: I integrate local real-time STT using Vosk/Whisper and high-fidelity TTS using Kokoro/Bark.
  • Quantization: I work with GGUF, AWQ, and llama.cpp to run large models on consumer hardware.
  • Analysis: I conduct token probability inspection and visualize attention maps to debug model behavior.