ආයුබෝවන් · I'm Nipuna Lakruwan.

I build software that thinks, and I measure whether it does.

Software engineer moving into AI engineering, from Kandana, Sri Lanka. Lately: an open benchmark of which embedding models understand Sinhala (now on PyPI), a Mac that complains out loud through a local LLM, and native macOS apps in SwiftUI.

  • Open to remote work
  • B.Sc. Information Technology
  • Sinhala · English
  • Powered by Ceylon tea

Click the laptop ·

Things I built (and debugged at 2 AM)

AI first. The 2023 C++ calculators are in witness protection.

ai · retrieval · benchmark

lanka-rag: which embedding model understands Sinhala?

Sinhala NLP resources are thin, so I measured. A reproducible retrieval benchmark of 7 local Ollama embedding models and 2 BM25 baselines on Sinhala, Tamil and Singlish (900 questions per task, with bootstrap confidence intervals), plus a local RAG assistant over 1,010 pages of Sri Lankan Acts.

  • bge-m3 leads Sinhala at 0.913 nDCG@10, but plain BM25 over character n-grams gets 0.899.
  • Python's \w regex splits Sinhala words apart, which quietly breaks many BM25 baselines.
  • Singlish is the open problem: the best model reaches only 0.31.
pip install lanka-rag

Plot twist: a 30-year-old keyword algorithm nearly beat a 568M-parameter model. I re-ran it three times.

  • Python
  • LangChain
  • Ollama
  • BM25 + RRF
  • PyPI
  • GitHub Actions

Sinhala retrieval quality

nDCG@10 on 900 questions · higher is better

Lighter bar: BM25 keyword search, no neural model.

ai · local llm · macos

Mac With Feelings

A menu-bar app that gives your Mac a mood, a face and a voice. It reads CPU, thermal state, battery, memory, disk and your browser tabs, then a local LLM writes what it says out loud. ("47 tabs? I'm literally sweating.") The little laptop on this page is its face.

  • Pure, tested mood engine (33 tests) separate from the hardware sensors.
  • Ollama for lines, a hand-written line book when it's offline, and a cleaner that strips thinking tags and emoji before speech.

Side effect: my Mac now judges my tab count. It is not wrong.

  • Swift 6
  • SwiftUI
  • Ollama
  • AVSpeech
  • IOKit · Mach
Nine laptop faces: thriving, chilling, sweating, hangry, overwhelmed, bloated, sleepy, worried and cramped

Native macOS

Android Mirror for macOS

A native SwiftUI front end for finding, connecting and controlling Android phones over ADB and scrcpy, with honest docs about what's verified.

Because squinting at the phone on my desk was beneath me.

  • SwiftUI
  • ADB
  • scrcpy

PulseEQ

A system-wide 10-band equalizer in the menu bar, with stereo meters, nine presets and automatic headroom so nothing clips. No drivers to install.

Built because my headphones deserved better than "Flat".

  • Swift
  • Core Audio
  • DSP

Visa Photo Guide

Crop passport and visa photos to official country rules, entirely in the browser. No uploads, no AI retouching, no libraries.

Smiling is between you and the embassy.

  • JavaScript
  • Canvas
  • Privacy

DevOps, data and full-stack

Weather Sentinel

Checks the rain between home and campus every 4 hours and pings Telegram only when it matters. Dockerized, scheduled by GitHub Actions, no server.

Sri Lankan rain doesn't send calendar invites. Now it sends Telegram messages.

  • Python
  • Docker
  • Actions cron

Lanka News Dashboard

Two services on a private Docker network that show, side by side, what a Redis cache hit saves over a miss.

Essentially a very elaborate way of saying "cache it".

  • Redis
  • Compose
  • Streamlit

Face-recognition attendance

Marks student attendance in real time from a webcam. Recognises several faces at once, with per-person cooldowns so nobody is counted twice.

The natural predator of proxy attendance.

  • Python
  • face_recognition
  • OpenCV

Sri Lankan tourism recommender

Recommends places to stay from real Sri Lankan Airbnb listings, using engineered features and cosine similarity on Spark.

Like a friend who knows every Airbnb in Ella, if that friend ran on Spark.

  • PySpark
  • Cosine similarity
  • Jupyter

PulsePlus

A full-stack fitness platform with AI-generated workouts, a community feed and a store.

Generates workouts. Does not do them for you. Yet.

  • React
  • Express
  • MongoDB

Selenium web-form automation

Data-driven UI tests with the Page Object Model, TestNG, Excel test data and failure screenshots.

A robot that fills in forms so I never have to.

  • Java
  • Selenium
  • TestNG
67★

Java Project-Based Learning, my most-starred repo: a path of progressively harder Java projects for beginners. Teaching is how I check that I really understand something.

Numbers nobody asked for

All true. Mostly.

47browser tabs open while building this site
4pieces Python's regex chops ජනාධිපතිවරයා into
3%free disk space on my Mac. Please send SSDs.
0frameworks used on this page
∞cups of Ceylon tea per benchmark run
1laptop with feelings, currently watching you scroll

The path so far

From cout << "Hello" to retrieval benchmarks, with only mild existential crises.

  1. 2023–24

    Fundamentals

    Dozens of small C++, Java and JavaScript programs: games, calculators, a banking system that, thankfully, never held real money. Plus a Java course people still star.

  2. 2025

    Full-stack and ML coursework

    React, Laravel and Express apps, Selenium test suites, a PySpark recommender and face-recognition attendance.

  3. Late 2025

    DevOps

    Docker, Compose, Redis and GitHub Actions: services that run themselves on a schedule, so I can sleep. (I don't, but I could.)

  4. 2026 · now

    Native apps and AI

    SwiftUI macOS apps, local LLMs, and measuring Sinhala retrieval instead of guessing. Also, a laptop with feelings. Long story.

What I work with

No 99% bars. Nobody is 99% at CSS. Not even CSS.

AI and retrieval

  • Python
  • LangChain
  • Ollama, local LLMs
  • Embeddings, BM25, hybrid search
  • Retrieval evaluation (nDCG, bootstrap CIs)

Apple platforms

  • Swift 6, SwiftUI
  • AppKit, menu-bar apps
  • Core Audio, AVFoundation
  • IOKit, Mach system stats
  • Swift Testing

DevOps

  • Docker, Compose
  • GitHub Actions, trusted publishing
  • Linux, Bash
  • Redis
  • uv, Ruff, pytest

Web and data

  • JavaScript, TypeScript
  • React, Next.js, Express
  • PHP, Laravel
  • MySQL, PostgreSQL, MongoDB
  • PySpark, pandas

Let's build something that thinks.

I'm looking for remote roles and projects in AI engineering, especially retrieval, local LLMs and tools for languages the big models overlook. I reply faster than my Mac boots.

nipuna@portfolio: ~

Prefer a terminal? Type help. Recruiters, try sudo hire nipuna.