Ilya Ivanov

I founded Heron inside Theona, where I'm Founding Product Manager. One question all along: how can a company trust AI agents with real work? The answer kept moving.

• Audit. We started by checking agents: an auditor that interviews an agent and compares its answers with its real configs and permissions. A one-time check, though, says little about what the agent does the next day.
• Gateway and observability. So we moved to runtime: one door between agents and company tools, with a record of every call. Every company needs that door, but it's table stakes, and a record shows what happened, not who answers for it.
• Governance. At dozens of agents the real questions are about the agents themselves: how many exist, who owns each one, what breaks when its builder leaves or a process changes.
• Rebuilding the process. Processes were built for people, so agents just produce more than the same people can review. We rebuild the process around the accountable person: they approve a rule once and see only the exceptions. Governance keeps it working as people and rules change.

How I work: most of my product work runs through agents, and the setup has changed every few months.

• Feb 2026 · OpenClaw. First personal assistant: Obsidian as a second brain, Fireflies call transcripts, monitoring of Telegram groups.
• Mar 2026 · A product brain in GitHub. Heron's strategy, customer research and call transcripts in one repo, so every agent (Claude Code, Codex) starts from the same context. Flow: Obsidian for raw thoughts, GitHub for shared context, Linear for execution, then code.
• May 2026 · A second brain and Alex. A personal vault built as an LLM wiki (raw sources, summaries, linked pages, a linter) next to the product brain: team knowledge in one, personal capture in the other. Alex, my chief-of-staff agent on top of Claude Code, was the Telegram front door to it: voice notes through local Whisper, calendar, capture.
• Since July · Claude Code directly. I dropped the Telegram layer. A session starts by loading my profile and the current state of work and ends with a close ritual that writes a session note and routes decisions, lessons and tasks to the right brain. Heavy research fans out to parallel subagents; long jobs run overnight against a validator. Theona agents handle lead research and outreach prep.
• Since August · Retros and shared rules. A weekly retro on how I work with agents: what held up, what broke, what to change. In October the rules moved out of local agent memory into the two repos, so every agent and teammate reads the same ones.

Automated mass creation of educational materials (2,000+ lessons). The agent takes a topic from Google Sheets, writes structured content with an LLM, formats a Google Doc, generates a PDF presentation via Gamma, organizes Google Drive, uploads the lessons to the learning platform and reports to Telegram.

Operational X-ray for companies in the chaos-of-growth phase: analyzes real workflows, finds bottlenecks and outputs a decision matrix. Paused.

Node-based editor that brings old family photos to life.

AI strategy for a fintech company: a 360° AI audit of its processes and two pilots, a multilingual support agent and a compliance agent.

Production Telegram Mini App with monetization: an AI companion with a defined personality and a 3D avatar.

Agent that scans subreddits over MCP for startup ideas and user pain points and aggregates what it finds.

Connected to my account, parses Bybit quotes in real time, decides with an LLM and trades futures on its own.

Launched an AI platform for de novo molecule generation. Cut molecule generation time from 2–3 weeks to 2–3 days and improved drug-likeness of generated compounds by 25% over classical methods. Built a 12-person cross-functional team from scratch and signed two R&D partnerships.

Grew from business analyst to head of the AI Lab in under three years. Closed a $400K pharmaceutical AI project, delivered a $150K industrial catalyst screening project, and built a 10+ person team from scratch.

Delivered two system modules end to end: specs, UI design, SQL, BPMN and UML flows, user support.