Lakshitha Kumaranayake Claim this profile

<!-- ───────────────────────── BANNER ─────────────────────────
banner.svg lives in this repo. No external service, no broken image. -->

<!-- ───────────────────────── ABOUT ───────────────────────── -->

ts
const lakshitha = {
  now:      ["SWE @ ELAROS 24/7", "Associate Lecturer @ Salford"],
  shipping: ["ENSO Watch", "LetterBridge", "ReMo", "MoRIS"],
  ai:       ["Claude API", "Genkit", "RAG",
             "structured output", "evals", "agents"],
  stack:    ["TypeScript", "Next.js", "Python",
             "FastAPI", "Postgres + Timescale"],
  domain:   ["Software development", "AI / Data", "health tech", "climate", "education"],
  status:   "Open to AI / full-stack roles",
  visa:     "Skilled Worker sponsor needed from 2027",
};

Ten years across engineering and teaching, in Sri Lanka and the UK. I work
where the output gets checked against reality — clinical assessment, benefits
letters, climate warnings — because that's where *"the model said so"* stops
being a good enough answer.

---

<!-- ───────────────────────── WORK ───────────────────────── -->

&nbsp; Systems in production

ENSO Watch

Real-time global El Niño / La Niña impact dashboard. Climate impact data exists but nobody can see it in one place. Open source, donation-funded, Sri Lanka as the anchor region.

FastAPI TimescaleDB Next.js Cloudflare

LetterBridge

NHS and DWP letters, in plain English. The people these letters are addressed to often can't read them. Decode, then answer the questions it raises.

Claude API Structured output AES-256 Stripe

ReMo

Physiotherapy adherence, measured at home. Clinicians can't tell whether home exercises were actually done. Pose estimation gives them something better than a self-report.

MediaPipe On-device inference NHS pathway

MoRIS

Autism and social-communication assessment. Slow, subjective, hard to standardise. Presented at the Cambridge Paediatric Health Tech Sprint, 2026.

Clinical tooling Child–parent profiling Research

---

<!-- ───────────────────────── PHILOSOPHY ───────────────────────── -->

&nbsp; How I build with LLMs

> The model is one component, not the architecture.
> Most of the work is retrieval quality, prompt versioning, structured output, and fallbacks
> for when the call fails. Knowing what to do when the answer is wrong *is* the job in health tech.

> Evals before intuition.
> A prompt that feels better and scores worse is worse. If a change can't be measured against a
> fixed set of cases, it isn't a change — it's a mood.

> Ship the boring parts properly.
> Encryption, audit trails, rate limits, cost ceilings. An AI feature that leaks a benefits letter
> is not an AI feature. It's an incident.

---

<!-- ───────────────────────── STACK ───────────────────────── -->

&nbsp; Stack

AI

Product

Platform

---

<!-- ───────────────────────── ELSEWHERE ───────────────────────── -->

&nbsp; Elsewhere

Teaching &amp; research

Associate Lecturer and Research Associate, University of Salford — Managing Information
Systems &amp; Transformation. Previously Senior Lecturer at ESOFT Metro Campus (2018–2023),
leading BCS and Pearson HND programmes.

MSc Data Science with Machine Learning, Distinction. Working towards SFHEA.

Before this

Digitalisation Support Officer, IOM — UN Migration. Social media coordination, FAO — UN.
Project Manager and Software Engineer, LK Technologies (2015–2021).

Companies

Director of Colombo IT (Pvt) Ltd (Sri Lanka) and KNOXIT Ltd (UK).

Most of the work above lives in private repos — health tech under NDA, mostly.
Working notes, updated as I build → lakshithalk.github.io

Collapse