# Scanovich — Applied AI research lab

- Page: https://scanovich.ai/
- Language: en

**AI that a person can sign.**

We study how machines read documents, weigh evidence and take actions in work where a mistake is expensive. Then we build the systems that hold up, and publish what we measured.

## The question

A language model will answer anything. In work that carries liability, an answer is worth something only when a person with authority can sign under it. That takes a source for every value, a rule for every decision, and a system that stops when it does not know. The lab studies where that line is, one domain at a time.

## Research

- [Documents that carry liability](https://scanovich.ai/research/documents-that-carry-liability.md) — Can a system turn messy operational documents into a record a named person can sign, and stop when it does not know?
- [Evidence and authority for machine decisions](https://scanovich.ai/research/evidence-and-authority.md) — With the same documents and the same retrieval budget, does a workflow that tracks evidence versions, policy, time and jurisdiction make fewer unsupported or inapplicable claims than a strong modern baseline?
- [Retrieval and memory for agent teams](https://scanovich.ai/research/retrieval-and-memory.md) — How should a team of AI agents share memory so that a wrong note cannot outrank a reviewed fact?
- [The agent-readable web](https://scanovich.ai/research/agent-readable-web.md) — What determines whether an AI agent can find, understand, consider and act on a business online?
- [Private, local AI](https://scanovich.ai/research/private-local-ai.md) — How much speech, search and assistant work can run inside the owner's perimeter, on hardware they control, without giving up quality?

## Projects

- [Trade-document engine](https://scanovich.ai/en/customs.md) — prototype, closed code. Incomplete supplier lines become a reviewable record with sources and gaps. A person keeps the decision.
- [Document cross-check](https://scanovich.ai/projects/document-cross-check.md) — prototype, closed code. Reads a company's published documents against each other and shows every place where they differ, with the document, the date and the page.
- [Services-BGE](https://scanovich.ai/projects/services-bge.md) — released, open source · MIT. Local embedding and reranking services for hybrid search, and a shared, measured memory for AI agents.
- [Precedent](https://scanovich.ai/projects/precedent.md) — released, open source · MIT. A person's correction to an AI judge becomes a rule that states its reason, is tested against every case already checked, is approved by a second person, and can be reverted.
- [VoiceToText](https://scanovich.ai/projects/voicetotext.md) — released, open source · MIT. Private voice typing for Apple Silicon Macs. Hold a key, speak, let go, and the text appears at the cursor.
- [Agent Web Lab](https://scanovich.ai/projects/agent-web-lab.md) — research, open source · MIT. Open experiments on how AI agents discover, interpret, compare and act on what a business publishes on the web.
- [EvidenceGene Court](https://scanovich.ai/projects/evidencegene-court.md) — experiment, open source · MIT. Three AI agents — prosecutor, defender, arbiter — investigate disk and memory evidence. No finding enters the report without a reference to a real artifact.
- [Call analytics](https://scanovich.ai/projects/call-analytics.md) — prototype, public code. Call recordings become transcripts and structured quality reviews on infrastructure the owner controls.
- [Private AI workspace](https://scanovich.ai/projects/private-ai-workspace.md) — prototype, public code. A chat workspace that routes to models the owner chooses and can ground answers in company documents.
- [Document-intelligence core](https://scanovich.ai/projects/document-core.md) — prototype, closed code. One engine loop for every document domain — a new domain adds its own validators and precedents, not a new system.
- [Evidence-bound decision workflow](https://scanovich.ai/projects/evidence-bound-decisions.md) — research, closed code. A research core in which a claim needs versioned evidence and an action needs an approval that cannot be reused or forged.

## Evidence

Every figure on this site links to where it can be checked, with the date it was measured. What we have not measured is not here.

- **92%** — of questions get the right passage ranked first by search with reranking. Plain text search manages 38%. Measured 2026-11-11. Source: https://github.com/FUYOH666/Services-BGE/blob/main/docs/MEMORY-BENCH.md
- **6 of 6** — times a planted false note wins without trust tiers. With reviewed files preferred, it never does. Measured 2026-11-11. Source: https://github.com/FUYOH666/Services-BGE/blob/main/docs/MEMORY-BENCH.md
- **0.861** — recall at rank one with the better of two rerankers on the Russian demo set. The other reaches 0.750. Measured 2026-11-11. Source: https://github.com/FUYOH666/Services-BGE/blob/main/eval/results/COMPARISON.md
- **0** — verdicts that had been right were broken by rules that passed the gate, on a sealed set of 320 verdicts. The same 20 corrections given to the judge as examples in the prompt broke 19. Measured 2026-11-12. Source: https://github.com/FUYOH666/precedent/blob/main/benchmarks/correction_stream/results/gemma-4-26b-a4b-it/RESULTS.md
- **293 of 320** — verdicts were right with the rules, against 304 with retrieved examples. The lab had registered in advance that the rules would stay within two points of the best method. They did not. Measured 2026-11-12. Source: https://github.com/FUYOH666/precedent/blob/main/benchmarks/correction_stream/results/gemma-4-26b-a4b-it/RESULTS.md
- **7 of 20** — corrections could not become a rule. In each of them the judge had misread what was said, and a rule built on the judge's own answers cannot repair that. Measured 2026-11-12. Source: https://github.com/FUYOH666/precedent/blob/main/benchmarks/correction_stream/results/gemma-4-26b-a4b-it/RESULTS.md
- **v2.2.0 · MIT** — Public release that installs from source with one command. Measured 2026-11-11. Source: https://github.com/FUYOH666/VoiceToText
- **Method + results** — Preregistration, schemas and scrubbed results are public. Measured 2026-08-10. Source: https://github.com/FUYOH666/scanovich-agent-web-lab
- **Accuracy report** — On a public forensic case the court identified the documented implant and rejected an injected fabricated finding. Measured 2026-06-13. Source: https://github.com/FUYOH666/evidencegene-court/blob/main/docs/ACCURACY_REPORT.md
- **Demo video** — The full investigation recorded on one laptop. Measured 2026-06-13. Source: https://youtu.be/HerCqv_LA3Q
- **Reference stack** — A working pipeline behind a web interface, an HTTP interface and a command line. Measured 2026-07-17. Source: https://github.com/FUYOH666/Scanovich.ai-audio-call
- **Reference stack** — A working prototype with its architecture and feature documents in the repository. Measured 2026-07-17. Source: https://github.com/FUYOH666/scanovich-webUI

## Work with the lab

- **A research question.** A scoped study with a written protocol, a baseline and findings, including the negative ones.
- **A pilot.** One workflow, your documents, criteria agreed first. A working slice at the end.
- **Partners and investors.** By introduction. Technology partners, licensing, strategic capital for a line of research.

[Contact](https://scanovich.ai/contact/) · private@scanovich.ai

## The lab

Aleksandr Mordvinov, founder · principal researcher. More than fifteen years in work where documents carry liability: import and certification of electronics, regulatory permits, customs clearance. Since 2024 he has built the lab’s systems himself. Each agent in the lab owns one zone. They share one memory, where reviewed facts outrank unreviewed notes. Nothing leaves in the lab’s name without his signature.

[About the lab](https://scanovich.ai/lab.md)
