AI Projects

Direct OpenAI (GPT-4o mini) and Groq LPU (LLaMA 3.3 70B) chat completions: system personas and prompts, temperature tuning, low-latency inference, and Pydantic schema validation.

LangChain Expression Language (LCEL) chains, multi-turn conversational memory, native OpenAI function calling execution loops, and LangGraph cyclical ReAct agent graphs.

Retrieval-Augmented Generation (RAG) architecture: recursive text chunking with overlap, 384-D dense vector embeddings (MiniLM), Chroma vector similarity search (cosine / k-NN), and context-grounded LLM generation.

Supervised machine learning regression (scikit-learn RandomForestRegressor), persistent vector database indexing and semantic retrieval (ChromaDB), and distributed LLM agent session memory (Redis).

Amazon Bedrock (Nova Lite Converse API), AWS Strands Agents SDK, LangGraph ReAct agents, dynamic tool schema reflection (@tool), stateful class tools, and async parallel tool orchestration.

Prompt engineering paradigms (Direct, Zero-Shot CoT, Few-Shot CoT), model alignment and sycophancy evaluation under adversarial pressure, and multi-model LLM-as-a-Judge jury consensus.

Chain-of-Thought (CoT) reasoning traces, Structured Outputs JSON Schema determinism, 2×2 tool schema routing calibration, and tool fault injection with hallucination prevention.

Model Context Protocol (MCP) client-server architecture via FastMCP, dynamic tool discovery and tool-calling agent loops, and automated tool routing and factual grounding evaluations.

Multi-agent orchestration architecture (Interviewer & Evaluator agents), ReAct reasoning and tool-calling execution loops, heuristic evaluation metrics, and shared conversational session memory.

LangChain Expression Language (LCEL) pipelines, LLM classification and conditional routing logic, policy-grounded prompt templates, and deterministic tool execution.