Independent engineering / working experiments

Code. Data.
Something you can try.

I build machine-learning systems from the dataset up. Explore the models, inspect what they learned from, and follow the engineering underneath.

Talk to my language model ↗Play the chess engine ↗

01 / LMSteinshark

A small model.
A complete system.

A 369-million-parameter language model: curated text, pretraining, instruction tuning, and cloud inference. Open the conversation first; inspect the architecture and training record when you want to go deeper.

PyTorchData pipelinesCloud inference

Open the playground →Code & architecture

MODEL / steinshark_369M369,167,360 parameters

32 transformer layers
16 query heads / 4 KV heads
30B pretraining tokens

→ Send a prompt. Inspect the response.

02 / cloudChess

179.9 million positions.
One move at a time.

From Lichess game records to a policy-value network and Monte Carlo tree search. Play the original bootstrap engine, then explore the larger dataset being prepared for the next models.

HDF5 / SQLiteResidual networksMCTS

Play & explore →Inspect the pipeline

DATASET / filtered Lichess2,331,246 retained games

Both players ≥ 2,000 Elo
Base time ≥ 300 seconds
179,935,641 position records

→ Filters → replay → tensors → training

The engineering is part of the work

Follow a record all the way to a prediction.

These projects include the parts between the headline numbers: filtering rules, reproducible datasets, tensor contracts, model serving, and tests that tell us when an experiment actually improved.

Language-model training monitor ↗Earlier project journal ↗

Outside the labProfile Services Finance Studio Ventures

Code. Data.Something you can try.

A small model.A complete system.

179.9 million positions.One move at a time.

Follow a record all the way to a prediction.

Code. Data.
Something you can try.

A small model.
A complete system.

179.9 million positions.
One move at a time.