Aquin is the research company reverse engineering intelligence with interpretability.
We build the tooling to debug and improve ML models by pin-pointing issues and simulating fixes, from reducing hallucinations to ensuring safety, based on mechanistic interpretability.
Backed by
build models with same precision as writing code.
Observe features and layers. Simulate training runs and catch failures. Diff base vs fine-tuned to see exactly what changed and why. Trace any output token back to the exact prompt span and layer where the answer formed.
Simulate, debug and improve ML models.
Inspect dense LLMs, MoE, vision, and embedding models. Trace any output token back to the exact prompt span and layer where the answer formed. Simulate LoRA, QLoRA, DPO, and distillation runs, catch failure modes before they compound, and diff base vs fine-tuned to see exactly what changed and why.
Interpret your existing infrastructure and pipelines.
Use the aquin watch CLI to observe external training runs. Run aquin watch init, ingest metrics JSONL with aquin watch ingest, and store runs locally under ~/.aquin/watch/. After aquin connect --name, loss, learning rate, and grad norm charts mirror to the web panel. Tail live with ingest --follow or replay with aquin watch on a run id.
The science: Mechanistic interpretability
Mechanistic interpretability reverse-engineers how neural networks compute, not just what they output. Aquin applies sparse autoencoders, logit lens, activation patching, and causal tracing to expose which features fire, which layers encode a concept, and which circuits produce each token. Stop guessing why a model hallucinated, drifted, or refused, trace the answer back to the exact prompt span that caused it and patch at the source.
Work with us
Interpretability tooling, custom SAE databases, mechanistic audits, circuit reports, and hands-on research, experiments, and studies for teams of all sizes. Reach us at aquin@aquin.app
Not sure if Aquin is right for you?
Aquin