Every role, audited like a pilot:
situation, execution, evidence.
login Walked Into
A blank page — and a conviction that non-technical people should be able to use AI. Built the company with three co-founders, later adding business-development and legal specialists.
construction Built
- ›Three product iterations, each to MVP: a pre-trained model marketplace, then "Adam" — a virtual AI engineer that inspects a dataset, cleans it, trains and deploys a model from natural language — then a full cloud interface.
- ›Learned the hard lesson: the bottleneck isn't the tech, it's that companies don't know which problems AI can solve. Pivoted to the "Show me → Help me → Do it for me" services framework.
- ›The discipline of killing my own weak projects — the same discipline the workshop enforces on yours.
task_alt Proof
- €50,000 angel investment secured.
- Google for Startups member — $350k in Google Cloud credits plus business-development consultancy.
- Graduated Founder Institute (2023) and Early Bird Vision Lab incubator (2024).
- Paying consultancy customer signed (end of 2024); technical proposal for Indora (global chemistry company) accepted on the merits.
- Pitched at Web Summit, Slush'D, and the European Business Angels Network Conference.
login Walked Into
First technical hire. The company had a basic unpublished PoC built by an outside boutique, and a founder-doctor personally serving pilot clients as a concierge MVP. No product, no team, no infrastructure.
construction Built
- ›Defined and shipped the MVP (medical questionnaires + telemedicine video calls) — which directly closed the first corporate customer.
- ›Architected the full preventive-health platform: daily health questionnaires, heart-rate-variability via phone camera, stress-hormone bio-sample tracking, plus a medical dashboard for specialists.
- ›Shipped an in-app AI assistant (live at departure), designed a recommendation engine for health pathways, and a prediction engine for health-trajectory interventions.
- ›Scaled infrastructure from two AWS VMs to a distributed microservices architecture on anonymised data (OVH). Hired and led a team of 10, training interns into full-time engineers.
task_alt Proof
- €1M corporate contract over 4 years closed on the platform.
- Designed for GDPR, EU MDR, EU AI Act, ISO 27001; independently evaluated as compliant from the product side — MDR Class II certification expected 2026.
- User base grown from a handful of test users to ~200 users.
- Accepted into the Future4Care accelerator (Paris); booths at VivaTech; brokered the OVH startup-program invitation.
login Walked Into
A post-acquisition mess: a customer-behaviour prediction engine that was Python 2 notebooks stitched together at 50–70% accuracy, owned by an acquired team in another city with a passive attitude toward the new owner.
construction Built
- ›Rebuilt the engine as Python 3 microservices with real-time inference — fixing data silos, data loss, and dirty local-database streaming along the way.
- ›Integrated results system-wide into CrossEngage's data stores instead of local storage.
- ›Founded a data science team and partnered on new models, replacing basic regression algorithms.
- ›Solved the human problem too: brokered trust between the acquired Hamburg team and Berlin management; was offered the Director of Engineering role (declined — for Sapiema).
task_alt Proof
- 99% uptime on the rebuilt engine; data-loss trigger points eliminated.
- A/B-scenario model predicting user action under competing campaigns with over 70% accuracy.
- Engine serving customers including Deutsche Bahn and Peek & Cloppenburg.
- Company successfully exited to Spotler.
login Walked Into
A company building an automatic recording, production, and streaming system for sports — with no automated production pipeline yet and, notably, no interview methodology or AI team structure.
construction Built
- ›From zero: trained object-detection and tracking models to follow the ball and every player across a 3-feed panoramic stitch, with a proprietary cluster-detection algorithm isolating the area of interest — processed at camera level.
- ›Solved event detection with a virtual field model reconstructed from visible line segments, placing players and ball in that space to track goals, fouls, and cards.
- ›For Deutsche Telekom: architected a home-streaming PoC training face- and voice-detection models in real time on the camera's own hardware, with pose-detection cropping and gesture controls.
- ›Built the team itself: job posts, interview methodology, hiring and onboarding of a senior AI engineer and a senior research data scientist.
task_alt Proof
- System deployed in production on a couple hundred sports fields, still evolving after departure.
- Real-time highlight detection at over 90% accuracy, generating tens of events per day for betting feeds and auto-produced match summaries.
- Telekom PoC implemented exactly as specified and accepted by the client (shelved only by a management decision to keep the product in-house).
login Walked Into
Amazon Marketplace content moderation at global scale: text moderation largely automated, but image moderation stuck at ~30% automation — and joined as a new-grad engineer with "theoretically limited responsibilities."
construction Built
- ›Personally initiated and brokered a collaboration with Amazon Robotics — outside day-to-day duties — to co-develop a comprehensive image-moderation model, then rewrote the image-moderation service with the squad.
- ›Co-founded a third squad building an org-wide model training-and-inference pipeline; defended it in front of the VP of Amazon Marketplace and got it approved.
- ›Re-architected legacy Java services into Scala on AWS microservices; key role migrating service communication from public internet to dedicated private TLS tunnels.
- ›When his original squad quit en masse, was asked to single-handedly keep its products alive and train the replacement squad — done; fully staffed and functioning at departure.
task_alt Proof
- Image-moderation automation raised from ~30% to ~80%.
- Pipeline handling millions of requests per day, mostly answered in milliseconds.
- Proposed for accelerated promotion to Senior SDE, skipping the mid-level entirely.
- Certified technical interviewer; mentored two junior engineers to promotion.
login Walked Into
The tools ad creators used to build ads on the platform — a company-level priority project on a four-to-five-person team.
construction Built
- ›Full-stack work on the new tweet creator for AdsEx — which became the backbone of the tweet creator in the platform's final pre-X version.
- ›Learned production full-stack engineering: REST services, React, JavaScript, Scala, Java.
task_alt Proof
- Advertiser engagement with the platform increased by over 30%.
- Release celebrated with a company-wide event; met Twitter's founders for a one-hour conversation on the back of the project's impact.
- Received a full-time offer — declined it to go deeper into AI engineering. A career bet on AI, made in 2017.
Before production systems:
first principles, in the lab.
Research at Jacobs University Bremen — where the habit of understanding models from the mathematics up was formed.
Bachelor Thesis — Neural Architectures
Six-month comparison of recurrent neural networks vs. echo state networks on protein secondary-structure prediction. Implemented the first attempt at training a bidirectional echo state network — proving it could match state-of-the-art performance with a fraction of the resources.
Array Databases for Satellites
Installed and optimised the rasdaman array-DBMS on an Arduino platform later used inside a satellite — making a full database engine run on radically constrained hardware.
MathWebSearch — Semantic AI
Worked on a search engine for mathematical formulas: integrated the GAP mathematical representation and guaranteed translatability into the group's semantic meta-language. AI from the symbolic, semantic-modelling perspective — the other half of the field.
Underwater Robot Control
Developed the engine-control system for a self-driving underwater robot — fusing autonomous routing with live operator guidance into a single control signal.
The paper trail.
Education
- BSc, Computer ScienceJacobs University Bremen · 2015–2018
- National High School of Computer Science "Tudor Vianu"Bucharest · 2011–2015
- Teaching AssistantComputer Networks · Data Visualization · C Lab — Jacobs University, 2017–2018
Honors & Awards
- INTEL ISEFInternational Science and Engineering Fair
- NASA Space Settlement Design Contestnational-level engineering competition
- Young Inventors Project Olympiadinvention & engineering olympiad
- Jacobs Hack 2015 · TechCrunch Hackathonhackathon awards
Skills, and where each one was earned.
Machine Learning & AI Engineering
Infrastructure & MLOps
Software Engineering
Leadership & Compliance
This is the experience behind
the Anti-Failure Engine.
Every framework in the AI Integration Workshop — the PRIME Filter, the CTO Veto, the pre-mortems — was paid for in production incidents, failed pivots, and systems that had to keep running. That's the point.
See the Workshop It Built →