Founder & Applied AI Architect · Scanovich
Production-grade AI infrastructure for sensitive business workflows — systems where data quality is imperfect, decisions are expensive, and trust is designed into the architecture.
Work spans:
- document intelligence
- voice and call operations
- retrieval-grounded applications
- governed multi-agent workflows
- local and private AI infrastructure
- compliance-aware automation
- forensic and audit-ready systems
Primary objective: risk reduction — data privacy, operational accuracy, compliance exposure, decision quality, and human sign-off when AI runs in production.
Public research explores one question from multiple domains:
How can AI systems operate safely around valuable data, sensitive workflows, and high-risk decisions?
Representative directions:
- agents that cannot publish unsupported claims
- retrieval systems with evidence-bound outputs
- supervisor-approved policy memory
- local-first call analytics
- audit trails and rollback for learned behavior
- private inference and retrieval infrastructure
- structured outputs for operational workflows
Domains differ; the underlying layer is consistent:
AI should not only be capable. It should be governable, observable, and accountable.
| Project | Focus | Repository |
|---|---|---|
| EvidenceGene Court | Autonomous DFIR with read-only MCP tools, adversarial review, fail-closed evidence gates, red-team harness, ablation, local model execution | evidencegene-court |
| ConductGene Swarm | Governed conduct QA: Prosecutor / Defender / Arbiter, Policy Genes, supervisor approval, audit export, Qdrant retrieval, live simulation | conductgene-swarm |
| AttestRWA | Settlement attestation and programmable escrow for real-world asset workflows | attestrwa |
| Scanovich Audio Call | Local-first call analytics: recordings → transcription → structured scoring → operational review | Scanovich.ai-audio-call |
Recommended profile pins:
evidencegene-court · conductgene-swarm · attestrwa · Scanovich.ai-audio-call
Local and private AI for sensitive data, internal knowledge bases, call recordings, operational documents, and regulated workflows.
Typical constraints: privacy, latency, data residency, auditability, model control, cost control, supervisor sign-off, integration with operator systems.
Extraction, classification, retrieval, and schema-bound outputs for noisy real-world documents — transport paperwork, customs-oriented inputs, supplier records, degraded scans, handwritten fields, declaration-ready structured outputs.
Speech-to-text, call analytics, QA scoring, and supervisor-governed review. Operational value: summaries, entities, risks, scores, policy references, and escalation signals — not transcription alone.
Multi-agent systems where autonomy is bounded by evidence, tools, approval paths, and audit logs.
Design patterns in active use: adversarial review, citation-bound decisions, abstain paths, human approval, rollback, deterministic baselines, red-team and ablation testing.
Autonomous forensic workflows with the model treated as untrusted by default — read-only tools, evidence re-derivation, tamper-evident logs, fail-closed publication gates for incident findings.
Document and intelligence systems for cross-border commerce, logistics, marketplace analytics, classification support, supplier normalization, and operational workflows.
Auditable release paths for real-world asset and settlement workflows — attestations, escrow logic, structured compliance evidence.
| Repository | Demonstrates |
|---|---|
| evidencegene-court | Autonomous DFIR court, read-only MCP, fail-closed validation, red-team harness, local execution |
| conductgene-swarm | Conduct QA, Policy Genes, supervisor-approved learning, audit trails, Qdrant retrieval, benchmark path |
| Repository | Demonstrates |
|---|---|
| Scanovich.ai-audio-call | End-to-end call analytics, structured scoring, saved results, pilot-ready API/UI |
| VoiceToText | Offline ASR, privacy-first speech processing |
| ai-agent-tts | Low-latency voice agents, streaming speech workflows |
| Repository | Demonstrates |
|---|---|
| Services-BGE | Embedding and reranking services for hybrid retrieval |
| linux-defender | Security-aware Linux operations, monitoring, audit support |
| Cleaner-OS | Workstation cleanup, dependency awareness, ML cache hygiene |
| Repository | Demonstrates |
|---|---|
| attestrwa | EAS attestations, programmable escrow for RWA settlement |
| realestate-agent-platform | Multi-channel enterprise agents, grounding, tenant isolation |
| Repository | Demonstrates |
|---|---|
| Scanovich.ai-MRI_radiology_assistant | Clinical imaging support, structured reporting research |
Languages and backend: Python, FastAPI, Node.js, TypeScript
AI systems: LLMs, RAG, agents, structured outputs, ASR/TTS, local inference
Inference: vLLM, llama.cpp, Ollama, LM Studio, OpenAI-compatible APIs
Retrieval: Qdrant, BGE embeddings, reranking, hybrid search
Agents and tools: MCP, LangGraph-style orchestration, tool-bound workflows
Data and operations: PostgreSQL, Redis, Docker, Linux, Apple Silicon
Quality layer: evals, health checks, audit logs, red-team harnesses, ablation tests, deterministic baselines
- Architecture before prompts — prompts guide behavior; architecture defines boundaries.
- Evidence before confidence — high-stakes systems require traceable evidence, not eloquence alone.
- Human sign-off where it matters — approval paths, audit trails, and rollback belong in the product.
- Private by default when data is valuable — deployment model matters as much as model choice.
- Determinism as a baseline — measurable behavior needs stable baselines, tests, and evaluation paths.
- Vertical slice first, hardening after proof — end-to-end paths ship early; production discipline follows demonstrated value.
Engagements typically align with:
- private AI infrastructure and enterprise adoption
- sensitive data workflows and call intelligence
- cross-border trade and marketplace operations
- governed agents and AI safety
- forensic and incident-response automation
- settlement, attestation, and audit-ready systems
- design partnerships and selective capital conversations across APAC
A substantive first conversation usually covers: workflow, data characteristics, risk, current manual process, decision point, deployment constraints, and human approval gate.
Email: iamfuyoh@gmail.com
Telegram: @ScanovichAI
Website: scanovich.ai
LinkedIn: Aleksandr Mordvinov