Projects
Research tools, production-shaped ML work, and app experiments.
Test-time training lab
A compact PyTorch lab for fast-weight and test-time-training ideas, with toy equivalence demos and a small language-model training harness.
active GitHub
Concensus SFT
A summarization modeling project with data preparation, fine-tuning artifacts, and loss-curve diagnostics.
paused GitHub
Second brain knowledge pipeline
A local-first knowledge system for turning papers and notes into reviewed article nodes, concept graphs, and publishable vault output.
active GitHub
Random Neighbors
Random-forest-style feature bagging for high-dimensional clustering experiments.
active GitHub
Reliability modeling templates
Production-style baseline templates for classification, survival, anomaly, and time-to-event modeling.
active GitHub
Field-to-test reliability modeling
Reliability modeling patterns for translating messy field behavior into test plans and failure-risk estimates.
active GitHub
Media ingest pipeline
A local-first ingest pipeline for transcripts, audio features, retrieval artifacts, and repeatable experiment runs.
active GitHub
Industrial anomaly detection
A defect-detection lab for reconstruction, segmentation, and evaluation on industrial visual anomaly data.
BirdCLEF audio modeling
Audio classification experiments around spectrograms, augmentation, validation discipline, and competition constraints.
active GitHub
Tabular foundation model lab
A practical lab notebook for TabPFN, TabICL, uncertainty, conformal prediction, and tabular baselines.
active GitHub
Agentic research template
A reusable repo shape for research agents: plans, artifacts, eval notes, and reproducible handoff state.
active GitHub
Local embeddings
A local embedding workspace for indexing, comparing, and inspecting text representations without external services.
active GitHub
Audio embeddings
Audio representation experiments across waveforms, spectra, model embeddings, and retrieval-ready artifacts.
active GitHub
Mixture-of-experts deep dive
A small lab for MoE routing, load balancing, expert specialization, and failure modes.
active GitHub
ResNet identity mini
A compact experiment around residual identity mappings, trainability, and small-model diagnostics.
active GitHub
RL context compaction
Experiments around using reinforcement learning to choose what context to retain, compress, or drop.
active GitHub
RL compaction
A smaller reinforcement-learning lab for compaction policies, rewards, and evaluation traces.
active GitHub
RL gym from Sutton
Small reinforcement-learning environments and experiments grounded in Sutton-style examples.
Streaming train demo
A demo of online training traces, incremental metrics, and model behavior while data arrives.
active GitHub
YouTube embedding pipeline
A local-first pipeline for audio download, Whisper transcription, text embeddings, and audio embeddings.
active GitHub
Complexity-aware program evolution
Program evolution experiments that track fitness, complexity, and the tradeoff between improvement and bloat.
active GitHub
Erdos concentration
An interactive app for concentration phenomena, tails, and the geometry of probability bounds.
Random matrix visualizer
An interactive visualization app for eigenvalue clouds, spectra, and random-matrix intuition.
Paxos explore
An interactive systems app for stepping through consensus messages, timing, and agreement behavior.
active GitHub
Tierra web
A browser playground for Tierra-style program evolution, mutation, replication, and population dynamics.
Blindwatchmaker
An interactive evolutionary-art playground for selection, mutation, and visual search.
LLM post-training harness
A small, explicit experiment harness for SFT/DPO and inference-time scaling (best-of-N, verifiers).
active GitHub
Vienna EP UI
A UI experiment around energy, inference, and controllable visual exploration.
active GitHub
Computational Life (ALife program soup)
A reproduction and extension of program-soup ALife experiments: BFF tapes, replicators, and phase-transition-like dynamics.
active GitHub
Event recommender system
A recommender-system project around event affinity, user behavior signals, and retrieval-ready recommendations.
paused GitHub
Opioid prescribing residuals
An applied modeling project for finding high-residual prescribing patterns after accounting for expected variation.
archived GitHub
Bayesian marketing attribution
Credible intervals and Bayesian regression for deciding which marketing channels are signal versus noise.
archived GitHub
Lead scoring ML system
An end-to-end lead scoring project across data cleaning, feature engineering, model validation, and deployment shape.
archived GitHub