A collection of agents, RAG systems, and ML pipelines I've built. Focusing on real-world utility and performance.
I built a personalized evaluation engine connecting technical dots beyond generic keyword matching. I designed a custom scoring context that understands experience depth.
I engineered a self-correcting agentic loop using Gemini 2.0 Flash. I implemented a 'Thought-Action-Observation' cycle, allowing the system to critique its own search results from Google Search API before synthesizing a final answer.
I developed a Multi-modal RAG assistant using Gemini 1.5 Pro Vision. I created a chunking strategy to preserve table structures and image context, ensuring strict context grounding to minimize hallucination.
I tackled the cold-start vs. idle-cost trade-off using a custom DQN Agent. I trained optimal scaling policies that outperform standard threshold-based autoscalers by anticipating demand surges.
I achieved <200ms latency on local hardware by pipelining Vosk, Ministral 3B, and Kokoro. I optimized the inference loop to stream tokens directly, creating a near-instant conversational experience.
I designed a semantic matching engine that bypasses keyword stuffing. I used Pinecone to store candidate embeddings and implemented a feedback loop to fine-tune the scoring algorithm.
I developed a background service that efficiently polls the Gmail API. I created a relevance filter that distinguishes urgency from noise based on sender history.
I bridged structured data with unstructured semantic search. I built a system that converts SQL product catalogs into vector embeddings, enabling natural language queries.