Writing & Research
Published notes only. Drafts and private research stay outside the public site.
Three-Stage Hybrid Search with Entity Routing
How to combine centroid search, parallel per-group HNSW queries, and entity routing into a recall-safe hybrid search pipeline.
Sleep Mode: Agentic Memory Consolidation
Designing a consolidation pass that merges similar concept groups using both vector similarity and LLM-based contradiction analysis.
Hierarchical Memory Retrieval with ChromaDB
How to structure persistent agent memory using concept groups and detailed cell retrieval in ChromaDB collections.
Building a Story Consistency Engine with Regex and Vector Search
Using regex-based fact extraction and vector-backed cross-referencing to detect contradictions in LLM-generated stories.
Local Agentic Tools with Ollama Function Calling
How to wire local LLM function calling into a modular tool framework, and what I learned building six tools for NeuroSavant.
Running a Vision LLM Locally for Real-Time Fitness Form Analysis
How the AI Fitness Trainer triggers Ollama + Qwen3-VL every 5 reps to deliver spoken form coaching — all on-device with no cloud dependencies.
Canvas City Map Rendering with Procedural Roads and Assets
Building an interactive 14x9 tile map renderer in plain HTML5 Canvas — procedural road connections, 40+ sprite variants, deterministic props, and full pan/zoom/click interaction.
Preventing Flicker with Dual-Arm Hysteresis in Bicep Curl Tracking
How a ±20° hysteresis bias stabilizes rep counting when both arms are visible during alternating bicep curl sets.
Building a Local Q-Learning Agent for Simulation AI
How I built a self-contained Q-learning policy engine for a city simulation — no cloud APIs, no GPUs, just numpy and a Q-table that fits on a Raspberry Pi.
What I learned training a 100M custom LLM
Key takeaways from training the Velora-100M transformer model from scratch, including data prep and learning loops.
Building a Local AI Fitness Coach with YOLOv11 and Ollama
Breaking down the keypoint extraction math, state machine rep counting, and local Vision LLM feedback loop behind a privacy-first fitness tracker.
Real-Time Face Recognition via WebGPU
Deploying YOLOv8 face detection and EdgeFace embeddings directly to browser GPUs for zero-latency client-side inference.