Fractional AI Lead

The high-stakes AI systems other teams get stuck on, shipped.

Machine Wisdom gives funded teams ongoing fractional AI leadership for complex AI and ML delivery. I embed as a durable technical lead inside the work: setting direction, making architecture and build-versus-buy calls, reviewing the work that matters, and creating continuity across architecture, delivery, evaluation, and technical decisions.

Book a 30-minute qualifying call See the engagements

Good fit

A funded team, complex AI/ML work, a thin principal bench, an internal owner, and budget for ongoing senior engineering leadership.

What I do

Embed as a fractional AI lead: set direction, make hard calls, write and review code, and unblock delivery.

Proof

Production ML at Roblox and Google; client safety architecture for a mental-health AI system.

Ways to Bring Me In

Start with ongoing fractional AI leadership when the work needs sustained principal direction. Use a sprint or review when the need is narrower.

Ongoing fractional engagement Fee confirmed after a qualifying call

Fractional AI Lead

You have complex AI/ML delivery work and want sustained AI leadership as your product, team, and systems evolve.

I provide principal-level direction without requiring you to build the whole leadership function in-house.
I make the build-versus-buy, sequencing, evaluation, architecture, and delivery calls that complex AI/ML work forces.
I create continuity across architecture, delivery, evaluation, and technical decisions while your team keeps ownership of the system.

Book a qualifying call

6 weeks Fee confirmed after a qualifying call

AI Initiative Recovery Sprint

You need concentrated principal attention on a defined AI/ML initiative, architecture decision, or delivery path.

I work inside the current product, codebase, data, and team constraints to separate the real delivery problem from noise.
I rebuild or reshape the highest-leverage part of the architecture alongside your engineers, so the work lands in your system.
You get a concrete path forward with owners, sequencing, and decision criteria your team can keep using.

Book a qualifying call

Focused review Fee confirmed after a qualifying call

Production Readiness Review

You need an experienced principal read on whether an AI/ML system is ready for the role the business needs it to play.

I review the product promise, system architecture, data path, model behavior, evaluation, and operational ownership together.
I identify the gaps that matter for delivery and distinguish launch blockers from ordinary engineering follow-through.
You get a clear principal recommendation on what to ship, improve, limit, or sequence next.

Book a qualifying call

Best fit when there is an owner who can act, a team already doing the work, and budget for ongoing senior engineering leadership. Fee confirmed on the qualifying call.

Proof Points

Selected work from production environments where technical choices had real operational consequences.

Client Engagement · Mental-Health AI

Safety architecture for a mental-health AI used by millions

For a major mental-health AI client, I served as embedded principal safety architect for a conversational AI companion. The design decouples safety from the conversation: classifiers score every message in full context, in parallel with response generation, and a safety gate decides pass, steer, or block before anything reaches the member. It introduced Empathetic Steering, redirecting unhealthy patterns without a hard refusal, on an append-only audit trail with shadow-mode validation.

Decoupled Safety ArchitectureEmpathetic SteeringAppend-Only Audit TrailMental Health AI

Read the case study

Embedded Principal · Disarray

Disarray

Disarray builds autonomous agents that turn complex proprietary data into production ML models, a frontier where the hard problem is not model execution but sustaining context, experiments, and institutional knowledge across long-running work. I worked with their team as a fractional principal engineer, contributing on two surfaces from a production-ML lens: the agent's memory and context persistence, and its evaluation, bringing the discipline of how you assess a model in production to a non-deterministic agent, separating real capability from reasoning, retrieval, and workflow failure. Disarray's performance on Kaggle ML challenges validates the technical core; the real work is making it hold in real enterprise environments. Engagement specifics are held under NDA.

Agent MemoryEvaluationAutonomous MLProduction ML Systems

Disarray About Disarray Doris Xin in The AI Journal Moustafa Abdelbaky in Unite.AI

From the Disarray blog: 1 2 3

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 DetectionNCMEC ReportingPatent-Pending

GitHub Roblox press release AP coverage

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 50% cost reduction and 300% latency improvement.

Ray Serve50k+ QPSCost OptimizationLatency Reduction

Read article

The same production judgment carries into fractional AI leadership: clear technical direction, grounded architecture, and systems your team keeps owning.

The Principal Behind Machine Wisdom AI

Younes Abouelnagah has spent years leading production ML systems where failure had real consequences: child-safety and abuse detection at Roblox, recommender systems and ranking at Google, safety architecture for a mental-health AI system, and agentic AI with teams building at the frontier.

The pattern is always the same: the strongest AI/ML work depends on connecting the model, the data, the evaluation, the product promise, and ownership into one system that holds together.

Machine Wisdom AI exists for those moments: principal-level engineering for teams that want ongoing AI leadership as part of how they build, operate, and evolve complex systems. The work is built around shared operating clarity, so your team keeps ownership while the hardest technical decisions get sustained senior attention.

Some work I don't take, on principle: AI companions and romantic products, weapons, and surveillance. The whole point of this craft is AI that helps people without harming them.

How engagements work View LinkedIn

Writing

Field notes from production AI

The same architecture judgment I bring to engagements, written down. Frameworks and post-mortems on how AI and automation systems move from tool choice to durable production architecture.

Quadrant map of the agentic automation landscape, plotting tools across problem type and governance maturity

Framework · Agentic Automation

The two axes that govern every agentic automation decision

The serious 2025 research, from MIT to McKinsey, BCG, and EY, converges on one shape: AI adoption is outrunning the integration, governance, and operating maturity that turn it into value, and the gap is organizational, not model quality. Two axes, the problem a system solves and the identity it is allowed to assume, map that gap into four quadrants and tell you what to buy, what to build, and what survives productionization.

Read the piece

More writing on AI safety, agents, and governed autonomy on Substack and LinkedIn.

Read on Substack Follow on LinkedIn

Focus Areas

The infrastructure gaps that decide whether an AI system can survive production traffic, enterprise scrutiny, and real user risk.

Identity, Permissioning & Audit

Production agents need to act as the right principal with the right scope, leave an immutable trail, and make it possible to reconstruct what happened after a consequential action.

Evaluation & Observability for Agents

Agent failures rarely have one cause. Evaluation has to separate model degradation, retrieval failure, reasoning breakdown, workflow mismatch, and tool-use errors while the system is still changing.

Safety Architecture in High-Trust Domains

Mental health, child safety, finance, and enterprise operations need policy, escalation, human review, and evidence handling designed into the system before the first serious incident.

Governed Autonomy

The hard move is not just a smarter model. It is moving from ungoverned automation to systems with identity, evaluation, audit, ownership, and rollback built into the operating model.

Book the qualifying call first.

The call is for fit: what is blocked, what deadline matters, who owns the decision, and whether principal-level embedded leadership is the right next move.

Book a 30-minute qualifying call

Not ready to book?

Send a short note with the initiative, what is blocked, the current deadline, and who owns the next decision.

Location

Toronto, Ontario, Canada
Serving North America, Middle East & Global

Connect with Younes on LinkedIn Machine Wisdom AI on LinkedIn

The high-stakes AI systems other teams get stuck on, shipped.

Ways to Bring Me In

Fractional AI Lead

AI Initiative Recovery Sprint

Production Readiness Review

Proof Points

Safety architecture for a mental-health AI used by millions

Disarray

Roblox Sentinel (SRIRACHA)

Fast ML Model Serving with Ray

The Principal Behind Machine Wisdom AI

Field notes from production AI

The two axes that govern every agentic automation decision

Focus Areas

Identity, Permissioning & Audit

Evaluation & Observability for Agents

Safety Architecture in High-Trust Domains

Governed Autonomy

Book the qualifying call first.

Not ready to book?

Location

LinkedIn

Send a Message