Trusted By 
We build interpretable AI models that replace the black box. You get a model whose reasoning you can read directly. Not a summary it wrote about itself. Not an approximation. A direct read of what it actually computed.
We work with teams in banking, financial services, and insurance where AI decisions carry real consequences and need to be understood.
Get started: pip install cobalt-ai
Organizations across banking, financial services, and insurance are deploying AI in lending, collections, claims, fraud detection, and compliance. The models perform. But when someone asks what actually drove a specific decision, the answer is either silence or a story the model told about itself.
Existing explanation tools were built for an older generation of models. They do not extend faithfully to the architectures being deployed today. As AI moves from pilot to production in high-stakes environments, this gap becomes the difference between a model that ships and one that stalls.
We work with your team to understand your current systems and build a model customized to your specific workflow. It bolts onto your existing infrastructure and is architected so its internal reasoning is directly readable. This is not post-hoc explanation. It is built into the model itself.
The model runs alongside your production system in shadow mode. No workflow disruption. No new integrations beyond a data feed. You get a side-by-side accuracy comparison on live data before making any production commitment. We prove improvement in your environment through a controlled comparison.
Cobalt is our interrogation platform. Business leaders can ask why the model recommended a specific action. Compliance teams can check consistency across similar cases. Technical teams can audit behavior at the level of internal representations. The answers come from what the model actually computed, not from a narrative it generated after the fact.
We partner with teams in banking, financial services, and insurance to bring interpretability to the AI workflows where it matters most.
When agents follow the AI recommendation most of the time, the model's reasoning needs to be documented. We provide the per-decision attribution record that risk and compliance teams need to see.
Automated lending decisions need transparent reasoning, particularly when built on non-traditional data. We surface patterns in model behavior that post-hoc tools miss.
The operational savings in fraud triage depend on scaling auto-close decisions. That requires a per-case attribution record at the point of disposition. We build the model that produces it.
When a model draws a conclusion from months of transaction history, you need confidence the conclusion was driven by the data. We make that reasoning visible and auditable.
Our approach combines two of the most rigorous methods for understanding AI systems:
TDA reveals the shape of high-dimensional model behavior without imposing assumptions. It surfaces clusters, transitions, and failure modes that standard evaluation misses.
We decompose model activations into interpretable features using sparse autoencoders and cross-layer transcoders. We map the circuits and concepts inside your model: not just what it predicts, but why, at the level of internal representations.
Together, these methods produce models whose reasoning is structurally readable. The result is interpretability that compliance and risk teams can verify, not just trust.
Cobalt is how your teams inspect, interrogate, and verify AI behavior. Built on topological data analysis and mechanistic decomposition, Cobalt gives business, compliance, and technical stakeholders direct access to a model's internal representations.
Ask a question about a decision. Get an answer grounded in what the model actually computed.
Install: pip install cobalt-ai
Inspect the internals of Qwen3 models. Trace circuits. Map concept evolution across layers. A free, interactive demonstration of an LLM's brain.
Our research drives what we build. Every method we publish is one we are actively applying to real problems with real clients.
Sachin Khanna
CEO
Gunnar Carlsson
Founder
Jakob Hansen
Head of Data Science
John Carlsson
Principal Scientist
David Fooshee
Principal Scientist
Founded by Dr Gunnar Carlsson, one of the inventors of Topological Data Analysis at Stanford. The founding team combines pioneering research in TDA and mechanistic interpretability with decades of enterprise software execution across global organizations.
Our advisory board brings deep BFSI credibility spanning tier-1 banking CTOs, AI governance leadership at global financial institutions, and PhD-level expertise in explanation-based AI. We navigate both the scientific complexity of interpretability and the operational reality of deploying AI in high-stakes environments.
