Loading stars...
Blackweb AI
Solving the frontier problems that decide whether AI can be trusted.
We build validation, inference, and control systems for AI operating in consequential environments — and we publish the evidence, including where it runs out.
Work with BLACKWEB AI
Have a broken integration or automation ticket?
Start with scoped engineering work, not a new AI platform. Diagnosis, reviewable repairs and regression tests—with scope and fee agreed before work begins.
See the engineering offer and send a ticket →KVRM / Research-stage · paid workflow evaluations
Choose the next supported action—not just the most likely one.
KVRM combines rules, retrieved cases, and compact learned models to choose among your operational actions. It filters unsupported primary candidates before ranking the remaining evidence, then validates the winner and any fallback. An inspectable Python decision engine—not another required LLM call.
One workflow. Five to ten defined actions. Sanitized historical cases. No production writes in the initial evaluation.
Compare against your existing rules and receive a written integration or no-go recommendation. A paid, fixed-scope evaluation. Price and timing are agreed in writing after scope and data readiness are confirmed.
Other systems and research previews
Separate systems. Explicit limits.
Both started as research and both run today. Each one publishes the number that would embarrass us if it were wrong, and each one tells you what it still cannot do.
oNeura
Simulate a drug from its molecular structure to a body-level readout.
72+ clinical trials replayed
Details and limits →LinguaGenesis
A language system that refuses instead of inventing an answer.
3,000 / 3,000 tampered proofs rejected
Details and limits →What we are working on
Open problems, stated as problems.
These are research directions, not products. Where we have results we publish them as papers, and where the results are narrower than the ambition we say so on the page.
Critical infrastructure AI
Systems that can propose a remediation are useful. Checking declared preconditions is one step toward deployment—not proof of real-world safety.
Hallucination-safe execution
A model can still be uncertain or wrong. The execution layer must be able to reject an unsupported action instead of turning a plausible answer into a real-world change.
Edge inference
We are researching how capable, bounded AI can operate nearer to the systems it controls, even when latency, connectivity, and auditability are constraints.
The important boundary
We do not claim to solve every hallucination problem.
Our validation work targets finite, governed action spaces: a runbook, a security playbook, an access-control decision, or a workflow with real consequences. It addresses structural checks around action validity; it does not replace judgement, monitoring, or a general-purpose language model.
oNeura is a research and strategic-IP program in computational biology, and our open-source work supports the wider Blackweb frontier-systems thesis.