Aquin

Aquin is best understood as a research lab rather than a single consumer app. The homepage describes the team as reverse engineering intelligence through interpretability, and frames its goal as applied prediction that compounds over time - systems that understand what a model has learned and that teams can point at their own data. The flagship applied product is Aleph, a universal predictor spanning tables, media, annotation and decisions.\n\nWhat separates Aleph from a general-purpose assistant is its output contract. It produces schema-bound results for programs: probabilities, labels, bounding boxes and scores, each carrying confidence so that uncertain items can be escalated rather than silently guessed. The lab positions this as 'one base, any modality, any task', fine-tuned into a specific workload and improving as more labelled work arrives.\n\nAround that core, Aquin publishes applied research on interpretability. Published work includes an interpreter for attention-as-routing-graph circuits from a single forward pass, tiled SVD methods for locating usable weight mechanisms, agentic weight-editing experiments on small models, and simulations of LLM behaviour in varied environments. Separate projects like the axi self-learning agent and the aq developer environment target people building and refining foundation models.