Notes · From the notebook
Half-formed thoughts,
honestly written.
No content calendar, no hot takes — just the occasional note about the craft, the journey, and what I'm learning as I go.
- Shipping the Claude-powered placement platform at Sansa Technologies.
- Starting my MSc in Data Science & AI at Moratuwa.
- Going deeper on evals and multi-agent orchestration.
- Plotting the next trip to somewhere with mountains.
●2026 · 4 min read
Why I left electrical engineering for AI
Four years of circuits taught me to love systems. AI is where I get to build the ones that think.
I spent four years studying electrical engineering in Islamabad, far from home. Somewhere between signal processing labs and my first YOLO model, I realised the thing I loved wasn't the hardware — it was the systems. The feedback loops, the way a small change ripples through everything.
AI is the same instinct at a different altitude. Instead of routing current, I'm routing information and intent — designing systems that read, reason and decide. The discipline of engineering never left; it just found a new medium.
So I made the jump deliberately. The electrical engineering degree isn't a detour behind me — it's the foundation everything else stands on.
●2026 · 3 min read
What 'forward-deployed' actually means
It's less about the model and more about the last mile — sitting with the client until the thing actually works.
Forward-deployed is a strange title until you've lived it. It means I don't hand off a model and walk away. I sit with the brokers, underwriters and carriers — the people whose day the software is supposed to change — and I stay until it does.
That last mile is where most AI quietly dies. The demo works; the deployment doesn't. Forward-deployed engineering is the commitment to close that gap: to own the prototype, the integration and the messy human edges all the way to production.
It's the most human part of a very technical job, and it's my favourite part.
●2025 · 3 min read
Teaching a model to read Sinhala & Tamil IDs
Most OCR is built for English. Sri Lanka runs on three languages — so the model had to as well.
National identity cards in Sri Lanka carry Sinhala and Tamil — two scripts that most off-the-shelf OCR quietly ignores. If you want to auto-register a company from an ID photo, 'quietly ignores' isn't good enough.
So we built OCR that reads all of it, and pipes the extracted director details straight into a business-registration platform through a custom API. No re-typing, no dropping the languages people actually use.
It's a small thing that says a lot: good AI meets people where they are — including in their own language.