Forward Deployed Engineer · Design Engineer · Bucharest · relocating anywhere for the right project
The agents write the code; I own the problem, the design, the tests and the review. Nineteen years in enterprise software, the last one at Matrix42, a German service-management and endpoint vendor, proving what one product engineer with coding agents can ship: a standalone AI product that became the company's AI, an MCP server that made a 20-year-old platform safe for agents, and designers who now ship production code.
The last three months · Sep – Nov 2026 · counted through 8 Nov 2026
- pull requests merged, Sep – Nov 2026
- median time from PR to merge
- commits, ten repositories, five teams
- colleague PRs reviewed
- tools in the MCP server I wrote
- production incidents from those merges
Act III
September. I stopped pairing with the agent and started running a system.
Plan, approve, build in a worktree, prove on a real tenant, open the PR. Ten repositories, five teams, no incident.
Featured · the last three months
One year of AI-first delivery, measured
Sep · Oct · Nov 2026
Bars on a square-root scale. Commits from git across twenty repositories, deduplicated; pull requests and merges from Azure DevOps. Counted through 8 Nov 2026 by a script that re-runs weekly; a hatched bar is the current month projected from its first days.
October 2026 · what I shipped
- Launch: Neo offered to 100+ enterprise tenants on 1 Oct; activation, consent and guidance flows.
- Nova program: 126 components, app-shell patterns, list view, data grid, graph canvas.
- AI Studio console spike across every page, then the blueprint for the port.
- 1,057 Claude Code sessions in the month.
- Commits
- 2,829
- PRs merged
- 97
- PRs opened
- 217
- Lines added
- 1.56M
- Test lines
- 355k
Let agents build
Every line I shipped in 2026 was written with Claude Code under my plan, tests and review: 160 pull requests merged across ten repositories, a median of 17 hours from PR to merge, no production incident.
Make contributing easier
esm-devkit turns a 20-year-old platform into something an agent can change safely: one write path with preview, read-back and rollback. 7,355 calls, no incident.
Review and security as automation
Local review before every PR, Sonar on push, 1,160 automated security-review runs. Reviewers commented on 9% of my PRs; the median merge took 17 hours.
Designers ship
I restarted the Nova design system as the shared repo and ran the upskilling program: Git, Claude Code, prototypes that become PRs. The first designer-turned-FDE now ships full-stack changes.
Act I
Three people built a standalone AI product. It became the company's AI.
August 2025 to April 2026. A founding team of three, outside the release train. Internal early adopters on regular calls from November; customers from January, shaping every release.
- my commits in the Cortex repository, #1 contributor
- of all commits came from the founding team of three
- lines I added, Jan–Aug 2026
- use cases in production by April
Cortex was Matrix42's standalone agentic AI app: chat that acts across the service desk, autonomous ticket resolution, root-cause analysis, knowledge management, routing and workflows, on web, desktop and Teams. I owned everything a user could see and a good part of what sat behind it, from the React 19 app and the Electron desktop to the marketing site.
In September 2026 the company made this technology Neo, the AI layer across its whole portfolio, superseding the earlier Matrix42 Intelligence features that larger teams had built over the previous two years.
- 150+ users and hundreds of interactions a day by April 2026, from a 5-minute self-service activation with no professional services.
- Generative UI spec: a render_ui tool so agents answer with interface, not prose.
- Eight languages, Electron desktop, Teams bot, ticket workbench with investigation summary and automation score.
- Demoed to the board, to Gartner for the Magic Quadrant, and at the European partner conference.
Act II
Then we put it inside a platform built in 2012.
May to August 2026. One chat component for every surface, consent flows an enterprise admin can accept, and a design system brought back from the dead.
- <m42-intelligence>: a Lit web component with React and AngularJS wrappers; streaming, human-in-the-loop tool approval, in-chat guides, account linking. I am its first contributor.
- Activation and consent flows written for risk-averse administrators: honest exits, a 30-day reminder, a forwardable draft, a data-protection path.
- Nova design system rebuilt to parity: 63 components in Lit and React, tokens, themes, Storybook, in one month.
- An AI-first experience whitepaper and a plan to rewrite the platform's frontend with agents, for the CPO.
Case studies
The products, in detail
Anonymised screenshots from test tenants and sample data; click any image to see it larger. Walkthroughs on request.
How I work
Six rules, learned by shipping inside a 20-year-old platform
Let agents build
Every line I shipped in 2026 was written with Claude Code under my plan, tests and review: 160 pull requests merged across ten repositories, a median of 17 hours from PR to merge, no production incident.
Make contributing easier
esm-devkit turns a 20-year-old platform into something an agent can change safely: one write path with preview, read-back and rollback. 7,355 calls, no incident.
Review and security as automation
Local review before every PR, Sonar on push, 1,160 automated security-review runs. Reviewers commented on 9% of my PRs; the median merge took 17 hours.
Designers ship
I restarted the Nova design system as the shared repo and ran the upskilling program: Git, Claude Code, prototypes that become PRs. The first designer-turned-FDE now ships full-stack changes.
Increase the ambition
Not issues, projects: a device-remediation app end to end, a design system to parity, a console redesign across every page.
Humans decide
Agents propose; people review, merge and deploy. I proposed the same rule for the whole company as an AI Coding Agents Message Board.
Before the agents
Designing and coding the interface since 2007
Meeting-intelligence startup. Designed and coded product and site: 6% landing conversion, 18% activation, $20k ARR in four months; led the company to the sale of its technology.
Microsoft Teams Approvals app (approval time −30%); first UX for the AI and cloud-first platform that became Cortex.
Anti-financial-crime and risk applications for analysts; global search for alert triage.
Designed and coded PayU's unified card checkout (still in production); Angular client apps at Teamnet; Neurons platform IA at Ivanti; UI kits and themes (#3 on Product Hunt).
I started in 2007 designing and coding interfaces, spent five years at Deutsche Bank on anti-financial-crime tooling, led design at Matrix42, co-founded an AI startup as CPO, and came back to Matrix42 in 2025 as part of the founding team of its Innovation Studio. In 2026 the job changed shape: I still decide how things look and behave, but the output is production code, shipped with coding agents, across repositories I had never touched before.
The right place is one where the product is agents in front of real customers, and one person is expected to carry a problem from discovery to deployment. Based in Bucharest, relocating anywhere for the right project; English and Romanian.
Built with Next.js, motion, a flow field and Claude Code. Every number on this site is counted, not estimated: commits, lines and test lines from git across twenty repositories, deduplicated by commit and with vendored bulk excluded; pull requests, merges and merge times from Azure DevOps. A script re-runs the count every week and republishes the page.