AI & machine learning research
Hatteria Labs studies how language models work from the inside. We publish what we find — including the results that failed — and build products on the parts that hold up.
Empirical work on model internals: memory compression, weight structure, quantisation, hybrid architectures. Every claim carries a null control it has to beat.
Methods, thresholds and negative results, in the open. A rejected hypothesis is a finding — it is where most of the value in this work actually sits.
Three products in production, built and run by the same two people who do the research. What we learn goes into software people pay for.
Research
Our current line of work asks how much of a transformer is redundant — and answers it by trying to remove things and measuring what breaks.
KV cache quantization is the only pain-free memory saving the project found for long context. A sweep over bits, granularity, and the K/V ratio on GPT-2 showed that keys need more bits than values. Porting the test to Qwen3.5 exposed a measurement-location artifact and showed the asymmetry is regime-dependent, not universal: keys win decisively below ~4 effective bits, while above ~5 the choice is domain-dependent.
GPT-2 small · Qwen3.5
A nine-step chain of tests asked whether GPT-2's transformer layers share enough parametric structure to store one common basis with a cheap per-layer description of the deviation. Sharing in raw weights did not hold; sharing in the functional space of input activations did, but only through fine-tuning, never through analytic projection. Attention matrices tolerate compression well (9.6x at a recoverable cost), while MLP layers structurally resist it.
GPT-2 small · GPT-2 medium
Products
Each product is live, paid for by its own users, and operated on our own infrastructure. They fund the research and keep it honest about what works outside a benchmark.
torumata.com
An audit for the era of AI search — see how language models actually read your site.
For: Site owners and marketing teams whose traffic is moving from search results to chatbot answers.
aidj.cloud
The DJ that runs your party on autopilot.
For: Weddings, company parties, bars and clubs, birthdays and school events — anywhere the music matters and nobody wants to babysit a playlist.
slidify.cloud
Your guests' photos, live on the screen.
For: Anyone hosting a wedding, a celebration or a company event who would rather collect the evening's photographs than chase them afterwards.
How we work
Most of what looks like a result in machine learning is a measurement artefact. Our process is built around catching our own.
We take on a small number of research and engineering engagements alongside our own products — mostly where a question needs measuring properly before anyone builds on the answer.
Get in touch