I build audit-grade AI systems — the kind whose output has to survive being taken apart by a lawyer, an auditor or an opposing party.
Multi-agent architectures, retrieval with citation-level provenance, evidence lifecycles for regulated operations.
Selected work
| SENTINEL | Adversarial audit council for police bodycam footage. Prosecution, Defense and Judge agents on deliberately different model families. Every verdict traceable to a specific utterance index and video timestamp, across four jurisdictions. Prize Winner, Milan AI Week 2026. Live demo |
| Claude Code delivery toolkit | Slash commands for client delivery — scaffolding, structured review, deploy verification, progress reporting. |
| Production systems | Lead-operations pipelines running 10,000+ executions for US design-build firms; multilingual voice-to-structured-event layers for field crews. Client work, mostly under NDA. |
Background — ran operations at a humanitarian organisation handling personal data of vulnerable populations under legal liability for data integrity: 1.5M+ PLN budget, zero audit findings. MSc Strategic Project Management, PRINCE2. Philosophy and history before that.
Currently — building FieldLedger, a multilingual evidence layer for UAE construction sites.
📍 karasov.co · v.karasov.pm@gmail.com