Open papers, careful methods.
We publish what we measure, and measure what we publish. Replication packets, evaluation harnesses, and weights are released alongside results.
2026
- N° 0052026.05.01field-note5 minKapllan Research
What a tokenizer is.
Why every model needs one, why the choices made today bind every model trained on top, and why we just spent ten hours getting ours right.
Read the essay → - N° 0042026.04.12essay3 minKapllan Research
The voice agent problem is not latency — it's turn-taking.
After three architectures and a year of patient defeat, every failure traced back to one assumption: that speech is a game of passes.
- N° 0032026.03.08essay4 minKapllan Research
Good search is not a retrieval problem — it is a calibration problem.
Every ranking signal we added improved our metrics. Then we added one verification step and discovered we had been confidently wrong about half the time.
- N° 0022026.02.21essay3 minKapllan Research
Bet on one model, own your fine-tunes.
We evaluated the obvious frontier candidates, and the choice came down to a single question that has very little to do with benchmarks.
- N° 0012026.01.22essay4 minKapllan Research
Software agents fail when they assume humans are patient.
Llana started as a code-generation tool. It became an autonomous coding agent because we restarted it fourteen times and read what was on fire each time.