Dmitry Ivanov

AI Implementation Engineer

I build Go-to-Market systems, AI-powered outreach, and automation pipelines that ship to production. From concept to deployed system.

Classification pipelines Cold outreach automation RAG & knowledge systems Multi-agent platforms
neurocraft — data classification reading text
> data classification _
01 / 05

From Idea to Production

I specialize in turning AI concepts into working production systems. My focus is on practical implementation: building RAG pipelines that handle complex documents, classification systems that process millions of records, and automation that replaces manual workflows.

Most projects in my portfolio run in production with real users and real data, while others are demos or proof-of-concepts ready to scale. I work across the full stack: from designing database schemas and vector search architectures to building React interfaces and deploying with Docker.

BasBot

Event-driven B2B AI agent platform: multi-LLM, multi-tenant, production-grade. Built on FastAPI + asyncpg + PostgreSQL with no LangChain, no Celery, no ORM. PostgreSQL is the only datastore and job broker.

BasBot runtime architecture Channels feed a host pipeline that reads memory from PostgreSQL and calls swappable LLM backends. A structured plan stops at an operator approval gate. Only then does a Docker-isolated executor run tools and return the answer. HOST SANDBOX IN PLAN ALLOW MEMORY MODEL IN Channels Telegram / Web PIPE Pipeline route · context · plan GATE Approval operator only BOX Executor Docker · no secrets DB PostgreSQL jobs · memory · graph LLM LLM adapters per-stage swap LEGEND input host gate / isolated store model

A plan cannot reach the sandbox until an operator approves it.

Multi-LLM routing
Claude, Gemini, Kimi, MiMo, GLM: backend + model configurable per-stage via DB. Swap providers at runtime without code changes.
Docker-isolated executor
Executor sidecar with iptables blocking private and metadata IPs. No host source mount. Credentialed ops go through named capability proxies.
Tiered memory + graph
L0/L1/L2 tiers with dual embeddings. BM25 + vector hybrid retrieval. Graph entity extraction. Context persists and compounds across sessions.
Operator approval gate
Planner produces structured plans; executor acts only after operator approval. Kanban board with real-time task lifecycle and artifact viewer.
Pluggable skills system
Runtime skill registry with per-bundle filtering. Skills declared in Markdown, injected at planning time. Zero hardcoded domain logic in core stages.
B2B multi-tenant
Workspace isolation, per-tenant LLM configs, browser session auth + Bearer API keys. FOR UPDATE SKIP LOCKED as the only job queue primitive.
Currently building

Company Brain

Durable, searchable memory for a whole company. It pulls in meetings, Slack, email, Jira, Confluence, CRM, ERP, and whatever else the team runs on, then turns it all into one typed knowledge graph where every fact keeps a link to its source. Ask a question, get an answer with citations back to where it came from.

company brain — evidence synthesis reading signals
> unstructured data
Slack “customers charged twice”
Meeting billing bug blocks release
Jira BILL-214 blocked for 6 days
HubSpot four related customer reports
Unstructured dataSlack, meetings, Jira, CRM

Every source stays attached to its original evidence.

StructureOne billing incident

People, reports, dates and decisions resolve into one graph.

InsightsRisk, pattern, next action

Cited conclusions become an operator-ready brief.

Evidence-backed synthesis
AI proposes topics, links and summaries; confidence gates, schemas and source provenance keep an uncertain conclusion from silently becoming a stated fact.
Entity resolution under uncertainty
Jira issues, meetings, people and customer requests get reconciled against each other, not just tagged. A typed schema blocks invalid graph shapes, but can't rule out a plausible-sounding relation that was never actually said.
Proactive gap detection
Finds repeated customer pain, stalled promises, and decisions that never became planned work: the problems nobody filed a ticket for.

Core Stack

AI / LLM

Claude OpenAI Gemini DeepSeek Kimi OpenRouter

AI Dev Tools

Claude Code Codex Cursor v0

Multimodal

Whisper ElevenLabs PaddleOCR

Backend

Python FastAPI asyncpg Node.js Rust C#

Data

PostgreSQL pgvector SQL / SSRS Power BI Redis

Infrastructure

Docker Nginx React TypeScript Alembic

Let's Build Something

Have an AI project that needs to ship to production? I'm available for contract work and consulting.