Curriculum vitae · print this page for a PDF
Bohdan Fesenko
I build the seam between AI agents and people, and then I make it reviewable.
Kyiv · remote · bohdan@fesenko.ai · github.com/bohdan-fesenko · fesenko.ai
Summary
Bohdan Fesenko is an engineer in Kyiv who builds systems where AI agents and people do real work together. He led engineering at a pre-MVP consumer startup, took a five-person team from no process to a shipped product, and made agent-assisted development a reviewable practice. He now builds agent platforms and measured detection-and-tracking pipelines.
How I work
Measure before you tune. The evaluation harness is written before the algorithm, and every phase is closed by a number on it, never by a feeling. Every parameter carries its provenance (measured, from a paper, derived, decided, or assumed) and the condition that would make me revisit it.
Name the failure mode before anyone asks. The strongest thing an engineer can do is state, unasked and with a number, the regime in which their own method stops working. My project notes carry falsifiers next to results.
Agents are part of the team, so their work is reviewed. I introduced agentic development to a company on day one and then built the rails: an onboarding course as a gate, review of AI sessions as part of code review, a scored rubric with named red flags. Velocity without reviewability is debt.
People get goals, not tasks. An engineer owns their module as its product owner. A tech lead keeps modules in sync, patterns consistent and delivery accountable, and is not a decision bottleneck.
Principles I hold
- Every phase is closed by a number on the eval harness, not by a feeling.
- One implementation, not a prototype and a rewrite: the harness measures the code that ships.
- Comparisons only at matched false alarms.
- Review of AI sessions is part of code review.
- Decisions are written next to the code, with the reason and the date.
- The only guaranteed loss, with zero variance, is the eternal brainstorm: never starting.
Experience
Engineering Lead
Pre-MVP consumer family app (stealth), COPPA-regulated
- Directly responsible for all engineering. Took a five-person team from no engineering process to a core loop live on TestFlight in about three months, then worked as platform architect and informal lead of the Ukrainian team.
- 148 merged pull requests across a seven-app monorepo in ten months; 95 database migrations, 452 test files and 29 CI workflows touched; about a third of the code volume while leading.
- Designed and shipped the product's entire MCP surface: an SSE resource substrate with a change-detection event bus, five live resources, an 83-tool catalog with per-tool annotations and boot-time registry validation, and OAuth with dynamic client registration for third-party MCP clients.
- Introduced agentic development on day one, then made it reviewable: an AI-coding course as onboarding gate, review of AI sessions as part of code review, a scored rubric with named red flags, an agent-orchestration work tracker, and a multi-model adversarial review tool.
- Wrote the team's planning and review protocol: goals not tasks, adversarial reviewers so knowledge spreads, approvals scaled to change size, review SLAs from four hours to thirty minutes for hotfixes.
AI engineer, agents and LLM applications
AI agents marketplace
- Built the agent backend and proof-of-concept agents for an AI agents marketplace, with a testing UI for evaluating agent behaviour before release.
- Shipped an LLM chat front end adapted to the marketplace's agents, and the demo the product was presented with.
Founder and AI lead
Own AI product venture
- Built generative-AI products with search and retrieval, agent workflows and custom tools.
- LLM applications for clients, including an ad-recommendation MVP with keyword analysis and several product front ends; a public Kaggle ship-detection solution.
AI engineer
Freelance client work
- Delivered more than 20 generative-AI projects for clients, from NLP chatbots to AI workflow automation.
Lead Machine Learning Engineer
ML services team, enterprise clients
- Led a team of three engineers: hiring, training and growth.
- Designed end-to-end ML solutions for enterprise clients, aimed at cost savings and process automation.
Machine Learning Engineer
Adaptive e-learning platform
- Built the machine learning pipelines behind personalised courses; course completion rose 23%.
- Built NLP answer-matching models that enabled new course features; bad reviews fell 37%.
Data Scientist
Computer vision company
- Built a face recognition system: a new network architecture, dataset and training method for image-quality estimation and embedding aggregation.
- Shipped the full pipeline as six neural networks on one server, serving four real-time Full HD streams on a single GTX 1080 Ti; deployed in several Ukrainian cities.
Selected projects
- track-lab: from noisy detections to confirmed tracks, measured — Targets too faint for any single frame, noise that looks like targets, labels that arrive late. A reference tracker and a simulator that knows the truth, with every decision measured and runnable in the page. Simulator with exact truth; Kalman and IMM; gating; GNN and JPDA; M-of-N and SPRT confirmation; track-before-detect; bearing-only fusion
- Agent-assisted development as a reviewable practice — I made AI sessions part of code review, with a rubric, named red flags and an onboarding gate. Velocity without reviewability is debt. Agentic workflow introduced to the company on day one; repo-resident agent commands and a story-sync workflow that made agent-authored story files in git the tracker of record
- An MCP surface that refuses to boot if a tool lies about itself — 83 tools, five live resources over server-sent events, OAuth with dynamic client registration for third-party clients, and a registry validated at start-up. 83-tool catalog with per-tool annotations and boot-time registry validation
- An assistant that operates this site in front of you — Ask it something and it flies to the page, points at the section, runs the demo and turns its knobs, then answers with cards built from the site's own content, so it cannot invent a number. One typed action registry drives four directors: the live model, recorded tours, voice, and any browser agent through WebMCP
- Few-pixel targets from a moving camera, in real time on the edge — A target three pixels wide carries almost no information in space. All of it is in time, so the pipeline integrates before it decides. Registration residual on real 4K moving-camera footage: held-out median 0.080 px over 706 pairs, zero failures, at the instrument floor
- One memory for a fleet of coding agents — Several agents, several machines, one shared data layer with mounts, signal streams, addressed whispers and a secrets vault, and a rule about who may write what. First generation: a single SQLite file exposed as an MCP server with 12 tools; documents with frontmatter, direct and broadcast signals, presence, full-text search; no external services
- A mobile product shipped to TestFlight in six weeks, by one engineer with agents — A social app for a city's motorcycle community, with a Telegram bot for the chats riders already use: iOS and Android from one codebase, one verification gate shared by people and CI. From the first commit on 2026-08-10 to a production build on TestFlight on 2026-09-17; 1,235 commits by 2026-09-23