I've run enterprise platform programs in regulated environments since 2018. At Medidata I grew a 50-engineer GitHub Copilot pilot into a governed GenAI program with 1,200+ users, and now chair the AI Operating Committee that decides which AI tools the company adopts. Nights and weekends, I build on-device AI apps and ship them to the App Store solo.
I run AI governance and enablement programs in regulated environments, alongside the enterprise platforms they depend on. The work spans program execution and platform operations: coordinating distributed technical teams, defining governance and workflow structure, and driving adoption for systems that make organizations more efficient and delivery more predictable.
Now at Medidata (Dassault Systèmes), leading enterprise platform programs, currently the Jira migration from Data Center to Cloud, alongside an expansion to Jira Product Discovery, Jira Service Management, and Confluence Cloud, paired with an org-wide ways-of-working rework. Led GitHub Copilot rollout across engineering and expanded a GenAI program to 1,200+ users. Coordinate distributed programs of 20–100 contributors with bi-weekly reporting to the SVP of R&D.
I hold PMP, CSPO, Leading SAFe, and Advanced CSM certifications.
I built both to run entirely on the device, and both started as a problem I had myself. I don't have an engineering degree and I've never held an engineering title: I write the code together with AI coding agents, the same class of tools I govern at work. It's the AI-enablement thesis tested on myself.
I fine-tuned a 246M-parameter grammar model, adapted it for the Apple Neural Engine, and shipped it as a native app: a correction takes about 17 ms, and nothing leaves the device. Apple's Writing Tools still skip Russian, so I built one that does.
Не могу дождатся выходных. Не могу дождаться выходных.
I built an Orthodox daily reader and trained a Core ML classifier for it that names an icon from a photo, on the phone; when it isn't sure it shows a shortlist instead of guessing. Alongside it I shipped the full Synodal Bible, a prayer book, and the liturgical calendar, with no accounts, no analytics, and one network request in the whole app, a date lookup for the calendar.
Icon recognition I measured at about 65% on clean shots · a guided crop lifted one test photo from 0.36 to 0.60 confidence · all on the phone
A browser tool, not an App Store product. I built an interactive body map that pinpoints exactly which muscle is spasming, down to deep ones like the piriformis, and gives the precise TENS/EMS pad placement, program, and session length for it.
The hospital gave me my MRI on a CD with a Windows viewer from 2009. I wrote down what finished looked like, pointed an agent at the disc, and left for the afternoon. It came back with a native Mac viewer, a from-scratch DICOM reader, and an MCP server so a model can look at the scan too. $30.80 in API-equivalent cost, 11,828 lines of Swift, and one installation guide that was almost entirely invented.
Eight years across three companies, the last five at Medidata (Dassault Systèmes), where I'm a Senior Technical Product & Program Manager. The resume is a download away at the bottom of this page. What it can't show is that the same five moves keep working, so here they are, each with the episode that taught it to me.
The AI Operating Committee approved everything until I limited submissions to department leads and moved tool costs onto the requesting department's own line. Twelve duplicate business cases surfaced before any money was committed; Finance and Operations had scoped the same problem with different vendors. The PMO intake got the same treatment: fewer recurring checkpoints, tighter gates, and processing time down 30%.
At Medidata, a 10,000-requests-per-minute API burst took Jira down and showed that nobody oversaw integrations. Instead of a post-mortem that faded, the Data Center migration got a steering committee, the platform's first API access policy, and response-time SLAs: an 8-hour cutover with no data loss, then 40% fewer ad-hoc support requests and 25% fewer integration incidents. At Symbridge the missing rule was access control. Security and Operations had deadlocked, so I proposed a hierarchical RBAC model aligned to NIST; admin-panel vulnerabilities fell 30% and SOC 1 and SOC 2 certification landed inside four months.
Three end-of-support deadlines (MySQL 5.7 to 8.0, 8.0 to 8.4, and MSSQL 2016 to 2022 across 20+ client-deployed Rave instances) carried $27M in extended-support exposure that R&D was quietly absorbing. Clearing the deadlines was the straightforward part. The work that mattered was moving what remained onto business owners' budgets, then building the Unified Ops Portal, about 15 operational tools folded into one platform with a financial-attribution layer, so the ownership stayed put. The same inventories became the foundation for FedRAMP Moderate readiness, and AWS rationalization plus a cloud cost-control policy added $20M in savings and $10M in cost avoidance over four years.
The GitHub Copilot pilot's success criteria were defect density and code churn, not licenses issued. On that evidence I argued for skipping the staged expansion leadership preferred; the rollout went to the whole engineering org, and the governance policy became the standard for every AI tool since. Non-technical teams came next on the same terms: an org-wide Gemini Gem and NotebookLM certification for 150+ people in Finance, Operations, and Marketing, which is how the program reached 1,200+ users. At Lumine the tool was an onboarding workflow and the outcome was new-hire ramp-up time, which it cut by about half.
Gramota's evaluation turned out to be scoring memorization: 22.6% on data the model had seen, 4.4% on data it hadn't. On messy real-world text the shipped tagger leaves 24% of sentences fully correct where doing nothing leaves 30%. Both numbers are on this site, because I'd rather publish them than pretend I never ran the eval. The MRI viewer an agent built for $30.80 came with an installation guide that was mostly invented; that's in the write-up too. Same thing at Lumine, years earlier: validating story sizes with engineering before sprint planning moved delivery predictability from 70% to 85%, and the number got reported either way.
Led a phased GenAI strategy from a 50-engineer GitHub Copilot pilot to an org-wide program reaching 1,200+ users: governance reform, an AI Operating Committee, and Gemini/NotebookLM adoption across non-technical teams.
EOL/EOS status, compute and storage utilization, and server inventory lived in five systems that did not agree with each other. I own the product work on the internal developer platform that puts them in one governed view; the financial layer is in development now.
Three database platforms reached end of support over four years, each carrying a seven- or eight-figure penalty for missing the deadline. I owned the programs that cleared all three (roughly $30M avoided), and changed the way the work was funded so teams prioritized it without being chased.
Managed the migration from Jira Server to Data Center, and used it as the catalyst to reframe Jira from a utility into a governed platform product with API policies, an SLA framework, and a steering committee.
The same review-protocol thinking I use chairing the AI Operating Committee, applied twice: a scorecard that sizes decision process to reversibility, and a 1944 sabotage field manual read in reverse as a management checklist.
Most bad decisions aren't wrong, they're mismatched: a week of review on something you could undo in an afternoon, or an afternoon spent on something you can't undo at all. This tool takes five answers about a decision and returns the protocol it deserves: who decides, how fast, and what has to be written down. A decision you can undo cheaply should not cost a week of meetings.
Describe a hypothetical AI use case in six clicks and get back the risk tier and the governance controls it needs under the EU AI Act and NIST AI RMF, with every number and citation behind the result shown. Runs entirely in your browser.
In 1944 the OSS published a field manual teaching citizen-saboteurs how to quietly wreck an organization: insist on channels, refer everything to committees, relitigate every decision. Declassified, it reads as a catalog of everyday management dysfunction. I reversed each instruction into a working principle for running programs.
Every tag opens the case study where the skill was actually used.
PMP (PMI, 2023) · CSPO (Scrum Alliance, 2024) · Leading SAFe 6.0 (Scaled Agile, 2025) · Advanced CSM (Scrum Alliance, 2025)