RL
post-training, distributed learning, and model analysis
- PyTorch
- TensorFlow
- GRPO
- PPO
- verl
- FSDP
- vLLM
- transformers
- sparse autoencoders
- scikit-learn
- reward engineering
- Qwen
Click the dark cells
Working stack
Agentic AI
tool-using systems with observable, verifiable behavior
- agent harnesses
- tool-use loops
- LLM evaluation
- verifiers
- MCP
- provenance
- in-context learning
- TabPFN
Molecular Simulation
simulation, stochastic systems, and chemical evidence
- NumPy
- Pandas
- PyTorch Geometric
- OpenMM
- MDTraj
- molecular dynamics
- Gaussian mixtures
- stochastic processes
- Monte Carlo
- computer vision
- spectral ML
System and Data Engineering
the infrastructure underneath training and live products
- Python
- C++
- SQL
- R
- TypeScript
- Docker
- Linux
- HPC
- DuckDB
- FastAPI
- SSE
- WebSocket
- tiered caching
Impact
Alkera AI
Data-agent task success
GRPO policy optimization across verifiable spreadsheet workflows
+340% improvement
4.4× higher10.5% → 46.2%
LangAlpha
Agent context overhead
Tool schemas compacted per LLM call; interaction latency also fell 140 ms → 40 ms
97% reduction
33× less10k tokens → ~300 tokens
SPEQTRO
Chemical-space exploration speed
Cross-modal candidate search across NMR, IR, and MS evidence
74.4% less time
3.9× faster2.6 min → 40 sec
Water · unsupervised ML
Ambiguous structural noise removed
Density denoising before the two-component mixture classifier
23% noise reduction
23% removedraw population → denoised population
Chem-ICL · Ersilia
TabPFN performance over Random Forest
Molecular-property prediction from a small labelled context
9.8% improvement
+9.8%TabPFN → RF baseline
What I'm learning
Working notes from classes, papers, and ideas I am still trying to make precise.
31 notes · 4 folders