An applied AI research lab. We train language models, engineer evaluation harnesses, and build vertically integrated infrastructure — because the most important advances in AI won't all come from the largest budgets.
Four threads of work, one conviction: the frontier is wider than the hyperscalers. Everything we learn here ships into our products.
We train and fine-tune specialized micro LLMs — small, task-focused models that rival general-purpose giants on the problems that matter, at a fraction of the compute.
We design, build, and stress-test agent harnesses and benchmarks. How models actually behave under real workloads is a research question — not an afterthought.
Reactor is our infrastructure research made real: an AI-native backend engineered end to end, from a single local node to a component-isolated fleet.
Compute optimization, local inference, and privacy-preserving architectures. AI that runs where the data lives — on your hardware, under your control.
Frontier AI is not reserved for frontier budgets.
The next wave of meaningful advances — compute efficiency, private AI, local inference, specialized models — will come from labs forced to be resourceful. We are one of them.
Nothing stays theoretical. Model training, harnesses, and infrastructure work ship directly into Super, AWEsome, Kani, and Reactor — and get hardened by real users.
We don't have infinite compute, and we treat that as an advantage. Working within tight budgets forces the optimizations that make AI cheaper, faster, and more private for everyone.
Benchmarks, whitepapers, and lab notes are shared openly on our blog — from AIBench evaluations to the Atomico PII Filter whitepaper.
From the lab
A bilingual DeBERTa-v3-xsmall encoder hits exact-span F1 0.897 English and 0.939 Spanish — above Grok 4.5 High on the same evals. The English-only predecessor did not clear the bar. Bilingual data did.
We evaluated eight frontier and budget LLMs across 240 machine-verified React and Rust coding trials. Quality still costs money, but retries, latency, and routing strategy matter just as much.
Kid profiles give each child their own encrypted space inside Kani — age-appropriate guardrails, separate conversations, and a Parent Dashboard that surfaces safety signals without reading every message.
