Writing about ML, data, and things I learn along the way.
True-dynamics CEM reaches the parking-v0 success diagnostic, then a scripted brake finishes the tape. Learned latent MPC parks 0 of 6 measured seeds. Residual, disagreement, and imagined spread sit at chance on crash-soon detection.
I froze the June WISDM notebook at git tag v1.0.0, then rebuilt subject-independent HAR. Same 5 s phone flatten: 0.8925 leaky macro-F1 vs 0.2924 GroupKFold. Watch statistical XGBoost is 0.7031.
A $15/month hobby lab around a thin FastAPI summarizer on one Hetzner box. Terraform, k3s, Helm, Argo CD, and Prometheus/Grafana/Loki are live. Autoscaling, secrets, TLS, canary, and disaster recovery are next.
How I built Gavel, a workspace for shaping ideas, collecting evidence, running structured multi-agent judgment, and turning model feedback into the next validation cycle.
Patterns and techniques for building RAG systems with LangGraph, Qdrant, and OpenAI. Covers chunking, hybrid retrieval, query rewriting, and hallucination prevention.
Building a bipartite genre network from the Steam database, projecting it into a weighted graph, running centrality and community detection, and learning why static game properties cannot predict playtime.
Implementing every ML algorithm twice: first from scratch in NumPy, then in PyTorch. From linear regression through CNNs, RNNs, and transformers, with miniprojects on MNIST, CIFAR-10, CelebA, and IMDB.