Deep applied ML experience
I have spent more than a decade building applied machine learning systems: recommender systems, search, ranking, trust and safety, and transformer-era model work at Google AI.
Applied ML systems leadership for production AI
I embed with your team and build production AI end to end.
Embedded applied AI & ML systems leadership: principal-level work across models, data, backend, frontend, and distributed systems. I've spent a decade on production large scale ML/AI including at Roblox and Google AI.
Good fit
You're committed to an AI mission and ready to let someone carry it with you. It can be a prototype or a legacy system, but ambitious and in need of a production-ready architecture that scales.
What I do
I work as your fractional AI lead, carrying decisions across models, data, backend, frontend, and distributed systems, from first exploration through scaled release.
Depth
A decade building production ML, safety architecture, and governed autonomy, from exploratory research to a system in front of users, accountable for what ships.
The hard part of AI work isn't the model, it's the middle: data, systems, identity, safety, cost, evaluation, release discipline, and the translation between engineering, product, and executives, all working together. That's where I live. I've built it from a blank page and scaled it once it mattered, and I carry the decisions either way.
Full-stack ML systems
The useful edge is not just modeling. I have shipped across data, ML, backend services, frontend surfaces, distributed systems, product constraints, and production operations.
Governed autonomy
The work I care about is AI that can act under identity, permissioning, evaluation, audit, rollback, and human ownership, rather than blind trust in a model.
Ways to bring me in
Ongoing fractional engagement
Fractional AI Lead
You need principal AI leadership embedded in the work as the product, team, and system evolve.
- Architecture, sequencing, evaluation, build-versus-buy: I turn ambiguous AI/ML direction into decisions your engineers can execute.
- I carry the hard technical calls and stay close enough to the code, data, and evaluation that they hold. And I bridge that deep engineering end to product and executives, translating the hard calls in both directions so the business understands what the technology can carry and the engineers understand what the business actually needs.
Time-boxed, usually about 6 weeks
AI Initiative Recovery Sprint
A consequential AI initiative is stuck or unclear, and you need a principal read that turns into movement.
- I separate the core delivery problem from noise inside your codebase and constraints.
- I reshape the highest-leverage part of the architecture alongside your engineers.
Focused review
Production Readiness Review
You want a principal read on whether AI/ML is the right tool for what your business is trying to do, and how to use it for that role.
- Product promise, architecture, model behavior, evaluation, and operational ownership reviewed together.
- Clear recommendation on what to ship, limit, pause, or sequence next.
This works best when stakeholders are committed to the mission, with a budget, and ready to give it room to move. We confirm scope and fee on the first call.
Selected work
Client Engagement · Mental-Health AI
Safety architecture for a mental-health AI used by millions
Embedded principal safety architect for a conversational AI companion. Safety decoupled from the conversation: classifiers score every message alongside generation, a gate decides pass, steer, or block, and Empathetic Steering redirects without hard refusal on an append-only audit trail.
Decoupled SafetyEmpathetic SteeringMental Health AI
Embedded Principal · Disarray
Disarray
Fractional principal engineer focusing on two surfaces: the agent's memory and context persistence, and evaluation discipline separating genuine capability from reasoning, retrieval, and workflow failure. Disarray builds autonomous agents that turn proprietary data into production ML models. Specifics under NDA.
Agent MemoryEvaluationAutonomous ML
Open Source & Patent
Roblox Sentinel (SRIRACHA)
At Roblox, I invented and led Sentinel, a contrastive-learning system for proactive detection of rare but severe harms. Public Roblox coverage reports 35% more actionable cases and over 1,200 NCMEC reports in the first six months.
Trust & SafetyReal-time DetectionPatent-Pending
Guest Publication
Fast ML Model Serving with Ray
At Roblox, I championed Ray for ML inference serving under high-throughput production constraints, with public Anyscale coverage describing a 50% cost reduction and 300% latency improvement.
Ray ServeCost OptimizationLatency Reduction
Frameworks, field notes, and architecture notes on production agents, governed autonomy, safety architecture, memory, and evaluation.
Read my articles on production AII'm Younes Abouelnagah. I go from exploratory research to systems in front of users, and I stay accountable for what ships. Machine Wisdom AI is how I bring that range to a team: embedded, hands on, and close to the decisions that matter.
Start with a 30-minute call.
The call is for fit: the mission, what's blocked, what deadline matters, and whether embedded principal-level leadership is the right next move.
Or send a short note: the mission, what's blocked, and the deadline that matters.
Toronto, Canada · Serving North America, Middle East & Global