edidiong umana · writing
rss study group

Writing

Notes from building, teaching and studying AI in Africa: agents that move real money, how to know they work, and a public log of my B.Sc. in Artificial Intelligence.

  1. Research notes10 min read

    The state of AI, and the next ten years: what I'm betting a degree on

    What is verifiable about AI in October 2026, six falsifiable bets for 2026 to 2036, what would change my mind on each, and what they mean for anyone training to be an AI engineer.
  2. Student-Builder Log · 004 min read

    I'm going back to school for AI

    I've joined the September 2026 cohort of the B.Sc. in Artificial Intelligence at Miva Open University. Why now, what I want from it, and how I'll study in public.
  3. Career playbooks6 min read

    What a forward deployed engineer actually does

    Three live forward deployed engineer postings, read line by line: what the role is, how it differs from solutions engineering, DevRel and software engineering, and how to build proof for each skill this year.
  4. Agents in the wild7 min read

    One run proves nothing: grading agents that move money

    Your payment agent passed the test. Run it eight more times. Why pass^k, end-state checks and invariants matter more than a demo that worked once.
  5. Engineering notes5 min read

    What an agent actually is: a loop with a model, tools, memory and a budget

    An agent is a short loop in your code, not a special kind of model. A runnable Python loop with hard step and cost limits, plus when to build a workflow instead, how to design tools and what MCP is.
  6. Engineering notes6 min read

    Start your evals from real conversations, not a benchmark

    Your AI feature passed every spot check and users still complain. How to read traces, name failure modes, count them, and build a small test set that looks like real life.
  7. Career playbooks5 min read

    Scoping an AI deployment before you write code

    Eight questions to answer before writing code when an organisation asks for an AI agent, drawn from Anthropic's and OpenAI's published guidance, with a one-page scoping template to copy.
  8. Engineering notes5 min read

    Can you trust an LLM judge? Check it against your own labels

    Your judge model says 80% of replies passed. How to test the judge like the classifier it is: TPR, TNR and kappa against human labels, its known biases, and paired comparisons that separate real gains from luck.
  9. Agentic commerce in Africa6 min read

    Mobile money already runs on agents

    African payments grew on human agents and savings circles. Five problems they already solved, from cash-out to reversals, and the design rule each one gives anyone building AI agent payments.
  10. Engineering notes6 min read

    The harness is the product: an agent is the model plus everything around it

    Your agent said it was done and left a broken app. A practical tour of the harness: AGENTS.md, skills, tools, permissions, sandboxes and checks, with a minimal AGENTS.md to copy.
  11. Agents in the wild6 min read

    Prompt injection is a design problem

    You can't stop a model being fooled, so design so that a fooled model can't do much. Least privilege, approvals, audit logs, canaries, safe rendering and red-teaming in every language you serve.
  12. Engineering notes6 min read

    Why your local model is slow: prefill, decode and the memory wall

    A long pause, then a slow trickle: two different problems with two different fixes. How to measure each one, and how to work out your laptop's decode speed limit with a pencil.
  13. Engineering notes4 min read

    Traces: what your AI system actually did in production

    Averages look healthy and users still wait. How to trace AI systems with spans and OpenTelemetry's GenAI names, keep personal data out, watch the few numbers that matter, and turn production failures into test cases.
  14. Agentic commerce in Africa6 min read

    Building AI under African data law

    You don't need to wait for an AI law. Data protection law in Nigeria, Kenya, South Africa and Ghana already governs your prompts, training data and servers abroad. The rules that bite AI builders, with sections and links, and a pre-launch checklist.
  15. Engineering notes5 min read

    Quantisation without the hype: what Q4 really costs you

    How to read Q4_K_M, Q8_0 and GGUF, size a model before you download it, and test the quality loss on your own task instead of trusting a leaderboard.