17-year-old developer from South Africa, specialising in AI agents, RAG systems, automations, and full-stack applications.
Not a list of skills. These are the actual problems I'm working through this month, and why they're the part that takes the thinking.
An agent runtime that can read mail, move calendar events, and edit docs on your behalf. It sits behind an approval layer, so nothing irreversible happens without a human saying yes. The interesting part isn't the model call. It's the permission boundary around it.
A RAG layer built for small knowledge bases where a wrong answer costs trust. It grounds every reply in retrieved passages and cites them, so you can check the source instead of taking the model's word for it. Right now I'm tuning chunking and re-ranking to cut hallucinations on messy internal docs.
Moving the workflows I first prototyped in n8n into plain code where the LLM makes the judgment call and typed functions do the actual work. Each step is testable on its own, so a change in one place can't quietly break the pipeline three steps down.
Each one solved a specific problem for real users. For each, here's what it fixed, how it's built, and the part that was genuinely hard to get right.
LiveTurns a customer's WhatsApp message into a professional quote in under a minute.
LiveA chat-based fitness coach that builds a real plan around your situation.
LiveA guided budgeting app that turns three questions into a money plan.
I keep the probabilistic part and the deterministic part separate: the model makes judgment calls, plain code does the work. Everything below follows from that.
Before I write anything, I want to see the actual job being done: the messy steps, the copy-paste, the thing everyone works around. Most of the real problem lives in details nobody thought to mention.
I sketch the data model and the boundaries between parts up front. Deciding what talks to what, and what's allowed to fail, is far cheaper on paper than halfway through the code.
I get one narrow path working end to end before widening it. A rough version you can actually click through tells you more in a day than a spec tells you in a week.
The first version is a question, not an answer. I put it in front of real users early and let what they actually do, rather than what they say, decide what changes next.
Getting it live is the start, not the finish. The errors, rate limits, timeouts, and monitoring are what separate a demo that works once from something you can depend on every day.
Understand → design → build → iterate → ship. Then harden it until it holds.
A running record of what shipped and what I learned along the way. The most recent work is at the top.
Paste a customer's WhatsApp message, get a branded quote back in under a minute. The hard part was pulling reliable job details out of messy free text, reliable enough to put a price in front of a paying customer.
A coaching chat that gathers your situation before it advises, plus an XP layer built to get people past week one. Learned that the engagement loop mattered as much as the model behind it.
Built Robotext a brand site with real-time 3D that still hits its performance budget. Most of the work was making WebGL behave on mid-range phones, not making it look good on my machine.
An agent that triages mail and schedules against a live calendar. This is where the approval-layer pattern I now reuse everywhere first earned its place.
Built an embedding-based search foundation so answers point back to real documents. This became the retrieval backbone under the assistants that came after it.
After a run of self-directed agent projects, I landed on giving the model a strict, typed tool schema and letting plain code do the work. Every agent I've built since starts from that boundary.
Wrote data pipelines and Puppeteer automations designed to retry and recover instead of falling over on the first timeout. Unglamorous, and exactly the part that decides whether a tool survives contact with the real web.
Live data pulled straight from GitHub: repositories, recent commits, and a year of activity.
Open to work · South Africa · Remote
If you're building something that needs an agent, a retrieval system, or a full-stack app around it, tell me what you're dealing with. I'll give you an honest read on whether I can help.
saadally038@gmail.com