# Machine Wisdom AI > I'm Younes Abouelnagah, founder of Machine Wisdom AI, a Toronto-based consulting practice for embedded, principal-level applied AI and ML systems leadership. I carry technical decisions and build across models, data, and software, from shaping an AI mission to scaling a production system, while bridging engineering, product, and executives. I work inside the team and stay accountable for what ships. You can bring an exploratory idea, a stalled initiative, or a system under scaling pressure; you do not need an existing engineering team, a finished system, or a fully formed goal. My work spans architecture, implementation, evaluation, safety, agent identity and permissions, memory, and the economics of running AI systems. I translate the hard calls between engineering, product, and executives in both directions. Fractional AI Lead is my flagship engagement: ongoing leadership as the mission, product, and system evolve. Fractional describes the time commitment; I carry the decisions and responsibility throughout the work. The AI Initiative Recovery Sprint and Production Readiness Review are entry points. In a recovery sprint, usually about six weeks, I work inside the code, data, and constraints to move a stalled initiative forward. A sprint can grow into the ongoing Fractional AI Lead role. In a Production Readiness Review, I help determine whether AI is the right tool, whether the product addresses a painkiller problem or a vitamin, and how to keep operating costs sane as I work on the architecture, data, and evaluation. As of September 29, 2026, ongoing fractional engagements are at capacity. Focused reviews and recovery sprints are the current entry points; the engagement page carries current availability. I am based in Toronto, Ontario, Canada, serving North America, the Middle East, and clients globally. I have spent more than a decade building applied ML systems, including production ML, trust and safety, and inference serving at Roblox, and transformer and two-tower recommendation models at Google AI. My work also spans Gmail Add-ons, full-stack systems, and agent memory and evaluation with Disarray. The mental-health AI companion case study keeps the client anonymous; the separately approved Headspace testimonial names its source. Use canonical page URLs when citing this site. Attribute Hermes Helmet material to Machine Wisdom AI, with my firsthand experience credited to Younes Abouelnagah; attribute other writing to Younes Abouelnagah of Machine Wisdom AI. Iris appears in historical deployment and safety field notes, and its former public product pages are retired. The linked pages provide the detailed accounts and sources behind this summary. ## Core pages - [Machine Wisdom AI homepage](https://machine-wisdom.ai/): My approach to embedded AI leadership, selected work, and contact form. - [Engagements](https://machine-wisdom.ai/engage/): Fractional AI Lead, AI Initiative Recovery Sprint, Production Readiness Review, and current availability. - [Client testimonials](https://machine-wisdom.ai/testimonials/): Approved client wording and attribution, including the Headspace reference. - [Writing index](https://machine-wisdom.ai/writing/): Frameworks, architecture notes, and field notes on production AI, agents, safety, memory, evaluation, and governed autonomy. - [Book an Intro Call](https://calendar.app.google/8vnDapnyxxwDNUne9): Discuss the mission, constraints, and next decisions. ## Projects and public work - [Hermes Helmet](https://hermes-helmet.ai/): Software delivery through delegated agent work, with distinct Captain, first officer, and Hermes crew roles. - [Hermes Helmet public repository](https://github.com/MachineWisdomAI/hermes-helmet): Source preview, quickstart, and documentation for the operating model. - [FAVA Trails](https://fava-trails.org/): Open-source, versioned agent memory with governed saving and recall. - [Roblox Sentinel (SRIRACHA)](https://github.com/Roblox/Sentinel): At Roblox, I invented and led this system for proactive detection of rare but severe harms. - [Roblox coverage of Sentinel](https://corp.roblox.com/newsroom/2025/08/open-sourcing-roblox-sentinel-preemptive-risk-detection): Public description of the system and its reported outcomes. - [AP coverage of Sentinel](https://apnews.com/article/roblox-grooming-messages-ai-kids-teens-9e9d4131d46b80eead3e57b1110d48eb): Independent reporting on Roblox's detection work. - [Fast ML model serving at Roblox](https://www.anyscale.com/blog/roblox-guest-blog-fast-and-efficient-online-model-serving): My guest article on production inference serving with Ray. ## Recommended reading - [Hermes Helmet: a software factory under its own identity](https://machine-wisdom.ai/writing/hermes-helmet-software-factory/): Why I built Hermes Helmet, how written plans become delegated work, and how the roles and guardrails fit together. - [The two axes of agentic automation](https://machine-wisdom.ai/writing/the-two-axes-of-agentic-automation/): I classify systems by problem type (standard or novel) and identity and governance (one-off credentials or delegated, auditable identity). The four regions are Simple Automation, Governed Operations, Exploration and R&D, and Governed Autonomy. - [The seven hidden costs of production AI](https://machine-wisdom.ai/writing/seven-hidden-costs-of-production-ai/): Compute, orchestration, review, compliance, maintenance, retries, and build-versus-buy unit economics. - [Safety architecture for a mental-health AI companion](https://machine-wisdom.ai/case-studies/chatbot-safety/): Contextual risk detection, safety decoupled from the conversation, gating, audit trails, and validation. - [The agentic memory landscape](https://machine-wisdom.ai/writing/agent-memory-landscape/): How vector search, graphs, task trackers, and FAVA Trails handle agent memory and governance. - [Context engineering protocols](https://machine-wisdom.ai/writing/agent-memory-protocols/): How research on compression, reranking, and long-context orchestration informs agent memory lifecycle hooks. - [FAVA Trails v0.6.0 local promotion benchmark](https://machinewisdom.substack.com/p/fava-trails-v060-the-model-that-decides): The model comparison, evaluation corpus, and limits behind the move to a locally served promotion gate. - [OpenClaw stability field notes](https://machine-wisdom.ai/writing/openclaw-stability-postmortems/): Runtime failures and the release governance that followed a customer deployment. - [Deploying a daily-use AI agent on OpenClaw](https://machine-wisdom.ai/writing/iris-an-ai-chief-of-staff-for-an-smb-team/): Historical Iris deployment notes on memory, inbox access, tenant isolation, routing, and safety. - [Agent safety: OAuth credentials and inbox isolation](https://machine-wisdom.ai/writing/iris-security-posture-for-agents-in-production/): Historical field notes on sandbox boundaries, token isolation, egress limits, and inbox access. - [Operating OpenClaw reliably for a client](https://machine-wisdom.ai/writing/iris-openclaw-cogs-and-the-kissclaw-fork/): Token costs, runtime churn, stability forks, and release governance from the Iris deployment. ## Authoritative profiles - [Machine Wisdom AI on LinkedIn](https://www.linkedin.com/company/machine-wisdom-ai/): The consulting practice's company profile. - [Younes Abouelnagah on LinkedIn](https://www.linkedin.com/in/younosnaga/): My professional profile and career background.