Open for projects · Nairobi & remote worldwide
I'm Mashel, an AI specialist in Nairobi, Kenya. I've trained frontier models for Turing, Outlier and iMerit, shipped agentic AI products, and taught 200+ people the craft, with a scientist's precision and dual degrees in Computer Science and Molecular Biology.
Trusted with frontier-AI training & evaluation
Most people in AI do one of these. The interesting work happens where all three meet.
RLHF, SFT, and multimodal evaluation for frontier AI. I grade outputs, build task trajectories, and catch the failures reward models miss.
Top 10% QA accuracy at Outlier AI · −25% annotation ambiguity at Turing
Agentic apps, RAG pipelines, and full-stack products. Function-calling, vector retrieval, and a frontend that makes AI feel trustworthy.
Onagi: a Gemini 3 vision + RAG agent with sub-millisecond reasoning latency
Workshops, cohorts, and one-on-one upskilling. I make the hard parts of AI legible to people who want to actually use them.
200+ students taught across 3 schools · mentored junior annotators
A few products I designed and built end to end. Each one starts with a real problem and ends with shipped, working software.
A marketing and lead-generation site for Roadbuck, a global supplier and manufacturer of automotive garage machinery, that turns a spec-heavy catalogue into quotes over WhatsApp.
An autonomous agent that watches the screen, reasons over what it sees with Gemini 3 Vision, and acts, writing to the database and calling APIs based on the user's intent.
A command center where small teams upload documents, get grounded AI summaries and drafts, and approve every automation before it runs.
A classic game rebuilt as a real product: an authenticated FastAPI + PostgreSQL backend, a 60 FPS React 19 frontend, and a CI/CD pipeline that tests and ships on every push.
Tell me what you're working on. Honest take, clear scope, no pressure.
One person across the modern AI stack, from training the model to shipping the product around it.
RAG systems, autonomous agents, function-calling pipelines, and AI features wired into your product or workflow.
RLHF/SFT data, model evaluation, red-teaming, and annotation pipelines, run with measurable QA discipline.
Production React, Next.js, and TypeScript. Dashboards, AI interfaces, and marketing sites that are fast and accessible.
Team upskilling, cohort-based masterclasses, and talks on practical AI, evaluation, and shipping with LLMs.
Landing pages, content, and the technical side of marketing for founders who need to be found and trusted.
Not sure which one you need?
Tell me the problem. I'll tell you honestly what would actually help.
Message meThree shapes the work usually takes. Price follows scope, so there are no fixed tiers, just a clear starting point.
| Project A defined build, scoped and shipped. | Ongoing A standing AI partner for your team. | Workshop Level up your team, fast. | |
|---|---|---|---|
| Scoped & priced upfront | |||
| Design & build | |||
| AI training & evaluation | add-on | ||
| Ongoing iteration & support | |||
| Team upskilling | optional | ||
| Async updates + scheduled calls | |||
| Start a project | Let's talk | Book a session |
We can all build safe superintelligence, one piece at a time.
I started in a molecular biology lab, where a 98% sample success rate isn't a brag, it's the price of being trusted. That training, evidence over vibes and precision under pressure, turned out to be exactly what modern AI is missing.
So I brought it to the models. I evaluate and train LLMs (RLHF, SFT, multimodal), catching the failures reward models miss. I build agentic products that use those models well. And I teach the craft, because the bottleneck on good AI is people who understand it, not GPUs.
The throughline is trust: making AI legible, grounded, and safe enough that a real person would stake a decision on it.
QA accuracy ranking at Outlier AI
annotation ambiguity reduced at Turing
people taught the craft
Computer Science + Molecular Biology
Notes on building AI products, design, and shipping. The thinking behind the work.
The hard part of an AI product isn't the model. It's making its output feel trustworthy and clear to a real person. A few patterns I keep reaching for.
Why I design and build instead of doing one and tossing it over the wall. What you keep when the same person owns both.
What I cut, what I kept, and how I kept a document-automation tool honest under a tight deadline.
The work is only half of it. Here's what I'm reading, building for fun, and thinking about when no one's paying me to.
I write about what I'm learning training and building with AI: practical, honest, no hype. One piece at a time.
The things people usually ask before reaching out. If yours isn't here, just send it.
AI features and agents, RAG systems, model training and evaluation, full-stack and frontend builds, and the occasional workshop. If it sits between 'understands the model' and 'ships a real product', it's my sweet spot.
Yes. I'm Nairobi-based and work remotely with teams worldwide. Most of my training work has been remote for US companies, so async collaboration across time zones is normal for me.
That's the whole point. I've done RLHF, SFT, and evaluation on frontier models, and I ship agentic apps and frontends. You get one person who understands both sides instead of a handoff that loses the plot.
Project-based, scoped and priced before any work starts. For ongoing work I'll set up a simple retainer. No vague hourly black holes; you know the number up front.
Usually within a week or two, depending on current commitments. Send me the rough shape of the work and I'll be honest about timing and whether I'm the right fit.
Yes, regularly. Conference talks, team workshops, and cohort-based sessions on training and building with AI. There's a dedicated page with topics and formats.
Open for projects · Nairobi & remote worldwide
An AI feature, an agent, a dashboard, a model-eval pipeline, or a team that wants to level up. Tell me what you're working on. I usually reply within a day.