I ship AI products — from idea to a running MVP.

Dvoechka

A full mini‑CRM for running school tests: regular and live sessions with a shared server timer, guest student entry, client + server autosave, anti‑cheat (copy/paste, tab‑switch, focus and DevTools detection), full session recording, quizzes with AI‑assisted generation and answer checking, and Telegram‑bot sync for teacher reports.

• React
• TypeScript
• Vite
• Supabase
• Edge Functions
• RLS
• rrweb
• Telegram Bot API

Legal Client Flow

A privacy‑first mini‑CRM for a lawyer, built fast. The focus wasn't CRUD — it was a real privacy/security layer: name/phone masking, click‑to‑reveal PII, copy guard, session‑only storage, explicit consent for external notifications and security headers. Plus webhook + Telegram automations. The same pattern — data masking, tiered access, audit-friendly logging — scales to business tools that need different security perimeters for different roles.

• React
• TypeScript
• Vite
• Netlify
• Telegram Bot API
• CSP

6‑month AI program · Academy TOP

A full educational product for adults — 192 academic hours, up to 30 participants. From research and course architecture to materials, practice, launch and feedback iterations. Covers LLMs (ChatGPT, Claude, Gemini, YandexGPT, GigaChat), n8n, local models, privacy, RAG and business automation, with ~15 applied mini‑projects.

• Curriculum design
• LLMs
• RAG
• n8n
• Automation
• Adult learning

PMD Stone

Designed a digital sales funnel: audience segmentation, competitive analysis, site audit, landing prototype, media plan, Yandex Direct & Avito channels, offers and an analytics loop — packaged as audit → launch → analytics → scale. Financial targets in the deck are goals, not achieved results.

• Positioning
• Funnels
• Landing pages
• Yandex Direct
• Analytics

Dual‑loop AI for a manufacturing plant

The client needed AI‑generated part instructions and reports, but part data could not leave the internal network. I designed a two‑loop architecture: an internal loop anonymizes real part data and generates randomized substitute data locally; an external loop runs the heavy AI models on that anonymized data; results then flow back into the internal loop, where the substitute values are swapped back for the real ones. The same pattern was used for report generation. No confidential part data ever reached an external AI provider.

• Data anonymization
• Dual-network architecture
• LLM orchestration
• Technical docs

Airlock: a private AI loop for documents

An open, testable version of the pattern from the confidential plant case above — plus the platform around it. The core (PII Airlock, open source): a local gateway finds sensitive values (deterministic rules + a local LLM), replaces them with opaque placeholders, calls the cloud only after a human confirms, and restores the real values locally. I measured what the AI actually adds: rules alone find 56% of the values, the union of two local models finds 92%. The platform (pilot-ready): an employee uploads a PDF, a scan, a DOCX or an XLSX, asks a question in plain language and gets an answer with links to the source fragments — behind a shared office password, with a warning if the same content was already uploaded under a different name. Under the hood: OCR, vector search with cross-encoder reranking on top, a local model, centralized logs and metrics, automated backups — all inside isolated Docker networks, with nothing leaving the machine by default. An ingested document is treated as data, never as commands: instructions hidden inside a file are executed neither by the model nor by the interface.

• 214 tests in CI (core + platform)
• 92% recall (model ensemble)
• 13 services in the loop
• 0 external calls by default

• Python
• FastAPI
• Docker Compose
• Qdrant
• Ollama
• Tesseract OCR
• RAG
• Reranking
• Local LLMs
• Prometheus
• Grafana
• Loki
• GitHub Actions
• pytest

Supplier search & procurement automation

For a tender/procurement team: AI searches for suppliers, fills a standardized template table and sorts it, with a manager doing a final manual check before it goes to leadership.

• 2× faster turnaround
• 3–7 → 1–3 days per procurement table + report

• Supplier search
• Data structuring
• Workflow automation
• Human-in-the-loop QA