Thoughts on ML, robotics, and engineering
9 entries · newest first
August 12, 2026
The zero from part 1 becomes 0.81 with a GRU bolted onto a frozen encoder, and 0.96 when the encoder learns from the prediction loss. Along the way: a model with the best one-step error in the experiment that leaves the data manifold by its thirtieth prediction, and an agent trained on nothing but hallucinations.
August 11, 2026
A VAE can reconstruct every frame perfectly and still know nothing about how the world moves. A toy environment where I know the ground truth, a linear probe that catches the latent out, a collapse that converged to exactly 11.17, and a dead model that scored 0.74.
May 2, 2026
How I built a pelvic fracture segmentation pipeline - and the single design decision that mattered more than the model. ABBC methodology, two-stage architecture with nnU-Net, and what actually matters when training on 3D medical data.
January 8, 2026
A walkthrough of building a transformer from scratch - RMSNorm vs LayerNorm, RoPE's rotation trick, the KV cache that makes generation not suck, and what training curves actually tell you.
December 15, 2025
The story of building BitTorch: ternary neural networks from paper to CUDA. Why my first kernel was 20× slower than PyTorch, the gradient bug that cost a week, and what I learned about low-precision ML.
May 24, 2025
System architecture development for the RC safety system. Covering functional architecture, physical architecture, and software implementation with ESP32 and Arduino, including state machines and failsafe mechanisms.
May 23, 2025
Diving into requirements engineering for an RC safety system. From stakeholder identification through mission definition to system requirements, following the V-model methodology with practical examples.
May 22, 2025
A prelude to a mini-series applying systems engineering principles to a 1/14-scale RC buggy project. Fresh from an AIAA course on space-grade SE practice, exploring the V-model methodology.
March 27, 2025
Starting a new series to share my learning journey through open-source courses like Geometric Deep Learning, Stanford CS149 Parallel Computing, CMU Optimal Control, and Applied Space System Engineering.