Deep-dives into AI architecture decisions, benchmark results, and engineering trade-offs from projects I've shipped.
I built MemoryLens after noticing that no publicly available framework quantified how different memory architectures degrade over extended conversations. Here's what I found.
How we built a full-stack industrial defect detection platform using DINOv2, OpenVINO, and a FastAPI streaming pipeline — and secured 4th National Rank at the Mitsubishi Electric Cup.
We replaced a $1,500/M GPT-3.5 pipeline with a $1.83/M dual-model ONNX ensemble and a 6-state FSM. This is how we did it.