00Prologue

The storyso far

I'm Lakindu Pahesara. Most people call me Pahesara. I build websites, apps and AI from Colombo, and I run a studio called CEYLAI Labs. Here is how that happened, in eight short chapters.

The short version

Paid client work since 2022. A data science degree, finished in 2026 with a Second Class Upper. Two AI systems built end to end. And a studio with eleven builds behind it.

Chapter 01 Colombo

An island with a useful clock.

Home is Colombo, on the west coast of Sri Lanka. The island is small, but its clock is well placed: at GMT+5:30, my afternoon is Europe's morning, and my evening is the start of New York's day.

That is how a studio in Sri Lanka works with clients in the EU, the UK and the US without anyone waiting a day for an answer.

A dotted world map with arcs from Colombo to the UK, the rest of Europe and the United States.
Colombo 
London 
Central Europe 
New York 

Chapter 02 2022

Paid from the first project.

I took my first paid build in 2022, while I was studying data science. Websites and apps at first, then the tools behind them.

Client work ran alongside the degree the whole way through, and each made the other better. Clients taught me that a model only matters once someone relies on it. The degree taught me to report results honestly, including the weak ones.

4 yrs

Paid client delivery before graduating

Early experiments, still on GitHub

  • AI Bookshop List Recognition System
  • Customer Churn Prediction ML App
  • AI Fitness Coach App
  • LLM Business Chatbot with RAG

Chapter 03 CEYLAI

A studio with one promise.

One person answers the client. That person is me.

CEYLAI Labs stands for Ceylon Artificial Intelligence and Digital Innovation Laboratories. A long name for a simple promise: one person answers the client, from the first call to handover. That person is me.

I win the work, set the standard and review every delivery before it ships, while the team builds to that standard. Eleven builds so far, for accountants, tour operators, exporters and a gem house in Ratnapura, plus ventures of our own.

  • The Northlight websiteNorthlight
  • The Reliance Business Partners websiteReliance Business Partners
  • The Ceylon Getaways websiteCeylon Getaways
  • The KYK Atelier websiteKYK Atelier
  • The Pandama.lk websitePandama.lk
  • The Serendib Exports websiteSerendib Exports
All eleven, and how the studio works

Chapter 04 2026

A degree, and two systems to show for it.

BSc (Hons) Data Science from the University of Plymouth, delivered through NSBM Green University. I finished in 2026 with a Second Class Upper.

My final year produced two AI systems, each built end to end. One reads brain scans. The other listens.

University of Plymouth

BSc (Hons) Data Science

Delivered through
NSBM Green University
Classification
Second Class Upper
Year
2026

Chapter 05 May 2026

In a stroke, the scan decides the treatment.

Treatment depends on knowing quickly whether a patient is bleeding or has a blocked vessel, and emergency imaging often waits for a radiologist.

I built a triage system that puts two models side by side, one on CT and one on MRI, and turns what they find into one decision a clinician can inspect: stroke type, severity, review priority and the evidence behind each.

On CT it reached a mean validation AUC of 0.9878 across six haemorrhage labels. The weaker MRI result is published too, with its spread.

A Grad-CAM heatmap over an axial CT slice of a brain, hottest where the model found evidence of bleeding.
Grad-CAM on a CT slice: where the model looked.

0.9878

Mean validation AUC across six CT haemorrhage labels

Read the case study

Chapter 06 June 2026

A chatbot that knows when to step aside.

The second system talks. It offers supportive conversation in Sinhala and English, from a Qwen2.5-7B model I fine-tuned with QLoRA and run locally.

Every message is read by a deterministic safety check first. High-risk messages never reach the model at all: they go straight to a crisis route with the 1926 helpline.

  1. 1 A message arrives, in Sinhala or English
  2. 2 A deterministic safety check reads it first
  3. 3Low risk: the fine-tuned model replies. High risk: a fixed crisis response with the 1926 helpline, and the model never sees it.
The Mental Chat interface with a message and a multi-sentence reply written in Sinhala. The risk indicator reads low.
A Sinhala conversation: translated in, answered, and translated back.
Read the case study

Chapter 07 Weekends

Off the clock, it is usually cars.

Lakindu Pahesara in a black suit leaning on a black BMW saloon, on a bright afternoon

Photographing them, modifying them, and building the tools that clean up the photos. One of those tools chains LaMa inpainting with rembg segmentation, so a car comes out clean with the distractions removed.

The portrait on my home page was cut out with a rembg model too.

Chapter 08 Now

Now, and what comes next.

Today I run CEYLAI Labs, roll the WhatsApp agent out to more businesses and keep two projects in the lab.

The next chapter is not written yet. It could be yours.

  1. Rolling out the WhatsApp agentTaking the multilingual commerce agent from Pehesara Book Shop to more Sri Lankan businesses as a CEYLAI service.
  2. Rebuilding ceylailabs.comThe studio's own site, rebuilt around its brand book: monochrome, with the drop as the hero.
  3. Shipping client sitesReliance Business Partners is live on a preview address, and the new Pehesara Engineers site is in draft.
  4. Two projects in the labOne is being built and one is being researched. Both stay under wraps for now.
Chapter 09 YoursStart the next chapter I reply within 24 hours, from me.