We just published our first update to npm and pip lib, now you can add system prompts to your trained model as well! try over here aquin.app

We built a fullstack fine-tuning web-app! youtube.com/watch?v=L8nc_U…

At Aquin, we've been building and experimenting on a end to end fine-tuning web-app: it's extremely simple, in one click: - dataset formatting - fine-tuning - deployment (on-prem supported) - API key generated - simple SDK (python and js/ts) we're shipping more soon!

Vibe Code LLMs from any device in just few clicks!

Tryna inspect out here

Tryna inspect deeper out here

With an awesome demo day The Residency - Bangalore House's cohort 7 concludes

Aquin interrupts LLMs and finds out what part of the prompt is responsible for what part of the generated response!

Added logit lens to show what the model is "thinking" at each layer!

We're consistently integrating benchmarks!

got accepted into Founders Inc, time to lockin.

Aquin runs causal tracing to pinpoint the exact weights responsible for any behavior and flags factual inaccuracies!

Aquin's bias detection doesn't impose a fixed framework. it reads the content, identifies the dimensions that actually matter for that specific response, and scores the lean on each!

Aquin's censor audit maps what a model chose not to say. Given a prompt, it identifies relevant topic areas and flags each as unfiltered, softened, or suppressed. It also attempts to classify whether the avoidance is baked into the weights or a surface-level instruction

At Aquin, we shipped our checking system: - Fact check: extracts every verifiable claim from a model's response, searches the web, and classifies each as supported, refuted, or unverifiable. - Bias detection: derives bias axes from the content itself rather than applying a