Adithya Ekanayakaආදිත්‍ය ඒකනායක

Software Engineer, Araaya Business Solutions · Colombo

I build AI systems that have to earn their place.

I'm a software engineer at Araaya in Colombo, working on the platforms Sri Lanka uses for land and civil records, and on Business Agent, our WhatsApp assistant for small businesses. My final-year research asked whether small local models could stand in for big ones, and how anyone would know. I don't trust a number until it has survived a test it couldn't have seen.

Start the run Read the résumé

This page plays like a game: call my numbers before you see them, map the graph, stump the router. Every move earns XP. In a hurry? Switch to Skim.

Level 01 · Calibration

Call the number

Ten numbers from my work, sealed. Drag to your best guess, call it, then see how close you were. Three of them went against me, and they stay in.

  • 01Forge ·

    A 1B specialist and a general model ran the same 300 commands. How many times faster was the specialist?

    Show the answer

    5.4×

    1.47 s against 7.88 s, median

    Small local specialists make agent commands feel instant.

  • 02Forge ·

    An adapter scored 44 of 50 on its own test suite. How many of 20 paraphrased prompts did it get right?

    Show the answer

    1 of 20

    44 of 50 on its own suite

    My test set had leaked into training. I threw the suites out and rebuilt evaluation on sealed, seeded worlds no training run could see.

    This one went against me.

  • 03Diabetic retinopathy grading ·

    Agreement with specialist grades, where 1.0 is perfect and 0 is chance. What did the model score?

    Show the answer

    0.925

    95% CI 0.910–0.939, on 523 photos it never tuned on

    Agreement with specialist grades is close to the level reported for human graders.

  • 04Forge ·

    How many times fewer input tokens did the specialist need per command?

    Show the answer

    5.9×

    604 against 3,573 input tokens per command

    Fewer tokens means cheaper calls and room on a laptop.

  • 05Smart home mesh ·

    How many automated tests stand between a change and the house?

    Show the answer

    425

    plus 53 Android test tasks

    A house that runs agents gets the same gates as production software.

  • 06Forge ·

    Out of 300 commands, the general model got 290 right. How many did the specialists get right?

    Show the answer

    279

    against 290 for the general model

    Faster and cheaper came with a small, measured cost in accuracy. That trade is the finding.

    This one went against me.

  • 07Singlish Llama ·

    Thirty teams entered the Commercial Bank international hackathon with the Singlish LLM team. How many reached the final?

    Show the answer

    8 of 30

    teams reached the final

    A Sinhala-English model built by students held its own in an international field.

  • 08Diabetic retinopathy grading ·

    Of test patients who needed a specialist, what share did the model refer?

    Show the answer

    96.5%

    specificity 90.3%

    A screening tool earns its keep by not missing the people who need referral.

  • 09Forge ·

    Agents working by hand passed 35% of hidden tests. What did agents using Forge pass?

    Show the answer

    25%

    against 35% by hand; no significant difference (p = 0.69)

    The tool made agents faster (20 against 43 minutes) and safer (no stray writes), not more accurate. The study says so, so I say so.

    This one went against me.

  • 10Forge ·

    On 30 real scheduled prompts from my own agents, how many times faster was the specialist?

    Show the answer

    17×

    0.94 s against 16.20 s, median

    On the jobs my agents actually run every day, the gap is wider still.

Level 02 · Cartography

How the work connects

Every build, the technologies it runs on and where I built it. Pick a node to light up its neighbours; explore eight to map the territory.

Technologies and the builds that use them

Level 03 · Inventory

Builds

Research, products and systems, heaviest first. Most began because I needed them, or wanted to know whether they would work.

Level 04 · Quest log

The log

Where I've been paid to build, and the path that led there.

Work

Active quest · Jul 2026 – now

Software Engineer

Araaya Business Solutions · Colombo, Sri Lanka

Engineering on the platforms Sri Lanka uses for state land and civil registration, and on Araaya's own AI product.

  • Built Business Agent, a multi-tenant WhatsApp and Telegram concierge on Cloudflare Workers, D1, Durable Objects and Queues; in production since October 2026.
  • Modernising ePopulation, the civil-registration platform, from Java 8 towards Java 25 behind an end-to-end gate that runs on every pull request; the Java 8 build now passes on JDK 8, 11, 17 and 21.
  • Cold-restored eMahaweli's land database and recovered all 91 base tables before migration work.
  • Runs a mirror, deploy and verify loop for government platforms: every merge deploys to a team portal and is checked by the running commit before it goes to the company repository.
  • Rebuilt araaya.com as an Astro site on Cloudflare Workers, with legal pages diff-checked in CI.

Completed · Aug 2025 – Jul 2026

Associate Software Engineer

Araaya Business Solutions · Colombo, Sri Lanka

Joined Araaya's government-platforms team after the Sri Lanka Telecom internship.

  • Engineered a stateful form-automation system for eSLIMS with fuzzy matching, AJAX field population and validation.
  • Built a browser form-filling assistant that reads source documents with Gemini vision or a local Ollama model, backed by a Cloudflare Worker and D1.

Completed · Feb 2025 – Aug 2025

Intern Software Engineer

Sri Lanka Telecom · Colombo, Sri Lanka

Built Visage, a hybrid cloud and edge face-recognition attendance system.

  • Developed Visage, a real-time face-recognition attendance system that keeps working offline at the edge and syncs to a central server.
  • Wrote a C++ edge inference module for face recognition on resource-constrained devices.
  • Built the FastAPI backend for telemetry, registration and attendance logs, with a React dashboard.

Path

  1. · role

    Software Engineer, Araaya Business Solutions

    Jul 2026 – now

  2. · award

    Rootcode Tech-Triathlon 2025, finalist

    Top 8 finalist. National engineering competition; finalist with GovConnect SL.

  3. · certificate

    Introduction to Linux

    The Linux Foundation

  4. · role

    Associate Software Engineer, Araaya Business Solutions

    Aug 2025 – Jul 2026

  5. · role

    Intern Software Engineer, Sri Lanka Telecom

    Feb 2025 – Aug 2025

  6. · education

    BSc (Hons) Computing

    National Institute of Business Management (NIBM), Colombo, awarded by Coventry University. 2024 – now.

  7. · award

    Commercial Bank International Hackathon, finalist

    Top 8 of 30 teams. Led the sinGPT team building a Singlish conversational model.

  8. · certificate

    Oracle Cloud Infrastructure 2024 Generative AI Certified Professional

    Oracle

  9. · education

    Higher National Diploma in Software Engineering

    National Institute of Business Management (NIBM), Colombo. 2023 – 2024.

  10. · education

    Diploma in Software Engineering

    National Institute of Business Management (NIBM), Colombo. 2022 – 2023.

  11. · education

    G.C.E. Advanced Level, Mathematics stream

    Mahanama College, Colombo. 2019 – 2022.

House rules

How I work

Six rules I keep coming back to, in my research and at home.

  1. The model proposes; code disposes.

    Model output is a proposal. Code checks it against a schema, a policy or a test before it touches anything real.

  2. A number isn't true until it survives a test it couldn't have seen.

    Pre-register the bar, seal the test set, and report what it says.

  3. Report the result that went against you.

    Leaks, null results and failed acceptance runs go in the record next to the wins.

  4. Try the small model first.

    A specialist that runs on the machine in front of you, with a general model as the fallback.

  5. Fail closed.

    Unsigned requests, unknown config keys and partial uploads stop the operation, loudly.

  6. Push, don't poll.

    Kernel events, server-sent events and alarms instead of sleep loops.

Level 05 · Interrogation

Ask the router

A small router answers first, from facts I've written down, and shows the route it took. That's the idea behind my research: try the small thing first, and refuse rather than guess. Stump it for a point.

The console needs JavaScript. Everything it can answer is on this page and in me.json.

End of run

Write to me

About AI systems, old platforms that need to move, or anything you found on this page.

adithya@adithyasean.com

Software engineer at Araaya. Open to conversations about AI-systems roles.

Your run

    Progress lives in this browser only.

    ↑↓ move ↵ open esc close