Machine-Learning

A comprehensive local RAG stack (ChromaDB + Ollama) designed for strictly private, reproducible retrieval and LLM inference; heavily focusing on hybrid retrieval strategies and index versioning.

A hands-on ML workshop that grew from a CLI script runner into a full-stack application: a FastAPI async backend dispatches training jobs to a thread pool, intercepts matplotlib output in-memory, and streams live telemetry to a Next.js dashboard via Server-Sent Events, while still supporting headless CLI experimentation with uv.

A comparative study of Apple-native frameworks (Vision, AVFoundation) against cross-platform C++/Python (OpenCV, pybind11) for video feature extraction.

An empirical systems exploration of local LLM inference on unified memory architectures: profiling compute-bound prefill vs. memory-bound decode phases, analyzing KV cache retention during cancellation, measuring the 64K vs. 128K context cliff on Apple Silicon, and comparing client payload strategies.

Guidance for reproducible, resource-aware machine learning experiments; leveraging lightweight MLOps primitives, strict environment versioning, seed management, and rigorous, deterministic model evaluation protocols.

Principles for highly available production ML inference; utilizing immutable model registries, dynamic request micro-batching, safe canary rollouts, graceful fallback degradation, and efficient GPU memory management.

Principles for architecting autonomous multi-agent systems; focusing on stateful orchestration, unified memory across agents, hand-off protocols, and human-in-the-loop governance for long-running workflows.

Best practices for resilient media feature extraction pipelines; ensuring stable representation schemas, choosing between streaming and batch modes, enforcing metadata preservation, and performance engineering.