2026
Research paper accepted at ISBM 2026; virtually presented in Bangkok.
hello, world —
Lakshitha Eranga
Systems Researcher & Architect
A boy from a small village in the Knuckles Mountains of Sri Lanka who grew up curious and reached for the sky and dreamt big. Today I’m a Computer Science graduate with First Class Honours from the University of Peradeniya and a Temporary Demonstrator there. I research the systems that run AI at scale and how to make them reliable and efficient at a lower cost.
Looking toward PhD research in systems / AI infrastructure
- C
- C++
- OpenMP / MPI
- CUDA
- Linux
- Python
- AWS
- Terraform
- Docker
- Kubernetes
Kandy, Sri Lanka
About me
Research, work & everything between.
pull any unit to read it
Research
2 papers1 in reviewMy interests sit at the intersection of systems engineering and machine learning, with a focus on AI infrastructure: making AI reliable, efficient and affordable. Watching distributed backends and asynchronous AI pipelines behave badly under load raised questions around resource management, scheduling, memory behaviour and fault tolerance. One of them, why parallelism stops paying off, is the subject of my ISBM 2026 paper.
Publications
- Accepted2026
Parallelism Isn't Free: Two Hidden Walls in a Heterogeneous Media-Understanding Pipeline
ISBM 2026, Bangkok — Springer, Lecture Notes in Networks and Systems
Two scaling walls found while profiling a production pipeline that mixes local CPU-bound models with external AI APIs. Virtually presented at ISBM 2026; proceedings to appear in Springer LNNS.
- Parallel Processing
- Concurrency
- Scheduling
- AI Pipelines
- Performance Analysis
This paper came directly out of production work on a heterogeneous media-understanding pipeline — one that runs local CPU-bound transcription models alongside external AI APIs for heavy video analysis. Scaling it revealed that adding workers stops buying throughput well before the hardware is saturated. The paper characterises two distinct walls behind that: one rooted in how contended CPU-bound stages interact with the scheduler, and one in how the pipeline's slowest heterogeneous stage bounds the whole graph. Both are measured against a real workload, and the paper closes with what this means for the cost of running AI infrastructure at scale.
- Under review2026
A two-stage output-space parameter-efficient transfer learning framework for low-resource medical imaging adaptation
Ceylon Journal of Science
A PEFT framework that adapts pretrained vision models in the output space, for medical imaging domains where labelled data is scarce.
- PEFT
- Vision Transformers
- Transfer Learning
- Medical Imaging
- PyTorch
This paper is the written-up result of my undergraduate research on transfer learning for medical imaging. Standard fine-tuning struggles when the target domain has only a few hundred labelled scans, and conventional parameter-efficient methods still reach into the backbone to adapt it. The framework proposed here works in two stages and confines adaptation to the output space instead, which keeps the pretrained representation intact while still fitting the new domain. It is evaluated against both full fine-tuning and standard PEFT baselines across several low-resource medical datasets.
Earlier research work
Optimizing Transfer Learning
University research · 2025
Researching PEFT methods and fine-tuning for optimizing transfer learning in low-resource medical imaging using Vision Transformers (ViTs).
- PyTorch
- Transformers
- ViT
- PEFT
- Fine-Tuning
This university research project explores the application of Parameter-Efficient Fine-Tuning (PEFT) methods on Vision Transformers (ViTs) for medical image classification. The focus is on achieving high accuracy in low-resource environments, comparing various PEFT techniques against traditional fine-tuning.
2026
Graduated with First Class Honours in Computer Science.
2026
Received the University Award for Academic Excellence.
2026
Earned Upwork Top Rated status.
2025
Organised REIGNITE ’25, reviving CSUP’s annual game night after seven years.
2025
Selected for the ICE Entrepreneurial Boot Camp, led by Prof. Suranga Nanayakkara; a 3-day programme with a 6% acceptance rate.
2024
Co-founded Xanvia and led development of the OpenEd LMS.
2022
Started a BSc in Computer Science and built the CSUP member portal, my first full-stack project.
2020
Started teaching myself Python, web development and Linux.
2015
Wrote my first “Hello World” in Pascal at Nugawela Central College.
BSc (Hons) Computer Science — First Class Honours
University of Peradeniya
2022 – 2026 · Minors: Mathematics, Statistics
Consistent academic excellence across all four years, alongside undergraduate research in transfer learning for medical imaging and leadership of the Computer Science Society.
Awards
- 2026
University Award for Academic Excellence · University of Peradeniya
Certificates
- 2021
Deep Learning Specialization · Coursera · DeepLearning.AI
Neural networks, hyperparameter tuning, structuring ML projects.
- 2022
Introduction to Big Data · Coursera · UC San Diego
Data management, processing and the Hadoop ecosystem.
- 2023
Google Cybersecurity Professional Certificate · Google
Security principles, network security, threat management.
Temporary Demonstrator
Dept. of Statistics and Computer Science, University of Peradeniya
Jun 2026 – Present · full-time
AI & Automation Engineer
Taino Spain S.L
2025 – Present · intern
- Built a multi-platform RAG chatbot (WhatsApp, Instagram, Web) on a single FastAPI backend, using LangChain and vector databases.
- Built an agentic social media content automation pipeline in n8n that generates, approves, schedules and posts content across platforms.
- FastAPI
- LangChain
- RAG
- VectorDB
- n8n
- Docker
Backend / Infrastructure Architect
contract
- Architected and shipped a production SaaS backend for AI video generation: event-driven services, pipeline orchestration, secure file storage, CI/CD and observability.
- Took an AI video job from ~10 minutes to a ~60-second pipeline — a 10x improvement in task execution time.
- Python
- Celery
- Redis
- AWS
- Terraform
- FastAPI
- Django
Full-Stack Developer
2023 – 2024 · intern
- Developed and tested features for client projects using Django (DRF) and React.
- Worked in Agile/Sprint cycles with Git best practices and Selenium UI testing.
- Django
- DRF
- React
- Selenium
- Git
Temporary Demonstrator
Dept. of Statistics and Computer Science, University of Peradeniya
Jun 2026 – Present
CSC 3081 · Undergraduate
Computer Graphics Programming
Practical lab sessions in C++ and OpenGL.
- AIWeb2026 · Lead Developer & Architect
SamajaAI — EdTech Voice Agent
Transformed an existing prototype into a scalable, production-ready MVP. Led the architectural decisions behind a secure EdTech platform built around AI voice agents, with ElevenLabs integration.
- AI Voice Agents
- ElevenLabs
- Architecture
- Security
- Scalability
Worked closely with the founder of SamajaAI to transition their working prototype into a robust, production-ready MVP. I made the critical architectural design choices that gave the platform long-term scalability and security. At its core sits a specialized AI voice agent tailored for the EdTech sector, leaning heavily on ElevenLabs integration for realistic, engaging learning interactions.
- AIWebSystems2025–2026 · Contract
Mixtape AI — Marketing Video Storytelling
Architected and shipped a production-grade SaaS backend: event-driven services, AI pipeline orchestration, secure file storage, and DevOps. Owned API design, job queues, background workers and the autoscaling infrastructure behind AI video generation — a 10x improvement in task execution time.
- Python
- FastAPI
- Django
- Celery
- AWS S3
- Terraform
Architected and built the core AI backend for a marketing-video SaaS platform, taking it from contract start through to production. The work covered integrating video generation models, orchestrating a hybrid pipeline that paired local CPU-bound transcription with external AI APIs for heavy analysis, running the asynchronous workload on Celery and message brokers so the main API stayed responsive, and managing scalable, cost-efficient storage on AWS S3. Infrastructure was provisioned with Terraform and deployed through CI/CD pipelines built for zero-downtime releases. The autoscaling redesign delivered a 10x improvement in task execution time.
- SystemsWeb2024
Self-hosted HLS Video Streaming Server (OpenEd LMS)
Problem · Mux/Vimeo costs add up fast for an LMS.
Adaptive HLS streaming with protection against easy downloads and full control over how videos are served.
- FFmpeg
- HLS
- Web2024 · Lead Developer — 3rd year project
OpenEd LMS (IT Center, UoP)
Led the end-to-end development of a comprehensive LMS for the University of Peradeniya, from planning to deployment. Features HLS video streaming and payments.
- Django
- TypeScript
- Next.js
- PostgreSQL
- CI/CD
- HLS
- Microservices
As the lead developer for our 3rd-year university project, I managed the entire development lifecycle. This comprehensive LMS, built for the university's IT center, features course management, secure HLS video streaming, content handling, multi-role authentication, and payment gateway integration. I was responsible for architecture, backend, CI/CD, and testing.
- AI2025 · Taino Spain S.L
Multi-Platform RAG Chatbot
A multi-platform RAG chatbot (WhatsApp, Instagram, Web) using FastAPI, LangChain, and vector databases.
- FastAPI
- LangChain
- RAG
- VectorDB
- Python
- Docker
This in-progress project involves architecting a sophisticated RAG system to handle queries from multiple sources. The goal is to create a unified AI assistant that can be accessed via WhatsApp, Instagram, and a web interface, all powered by a single FastAPI backend and a Graph RAG implementation for complex, contextual answers.
- AIWeb2025 · Upwork project
MyDresser — Virtual Try-On
Building a virtual try-on app with AI-powered features. Responsible for backend development and AI implementation, making the system production-ready, fixing bugs, and ensuring scalable development practices.
- Django
- Google Generative AI
- Python
Developing a virtual try-on application that leverages AI to provide a realistic fitting experience. As the backend engineer and AI developer, I'm building the AI features to power the virtual try-on functionality, ensuring the system is production-ready, fixing bugs, and implementing scalable development practices to support future growth.
- AI2024 · Xanvia
AI Trading Bot
Automated a UK client's binary options trading strategy, integrating a ChatGPT-based LLM for auto-trading decisions.
- Next.js
- ChatGPT
- OpenAI API
- Automation
This Xanvia project involved automating a client's complex trading strategy. I developed a NextJs application that interfaced with a binary options site, analyzed market data, and fed that data to a ChatGPT LLM for trade decisions. The system then executed trades automatically based on the LLM's response.
- Web2024 · Xanvia
studentmoves.co.uk
Co-led design, architecture, and development of a property rental service for UK students, including Stripe payment integration.
- Django
- Next.js
- Stripe
- PostgreSQL
- Tailwind CSS
As a co-founder and lead developer at Xanvia, I architected and built this platform from the ground up. My responsibilities included designing the database schema (PostgreSQL), developing the Django backend, and integrating the Stripe payment gateway. This project involved full-cycle SaaS development, from idea to deployment.
- AI2025 · Taino Spain S.L
Social Media Content Automation Pipeline
An agentic, multi-platform content automation pipeline using n8n to generate, manage, and distribute social media content.
- n8n
- AI Agents
- Webhooks
- REST APIs
- Python
This project involves building an intelligent workflow using n8n to automate the social media content lifecycle. It uses agentic AI to handle content generation and adaptation, while the n8n pipeline manages approvals, scheduling, and posting across multiple platforms (like Instagram, X, and LinkedIn) via their respective APIs.
- AI2025 · Upwork project
Multi-Agent UI Tester
Fixed and integrated a multi-agent system, connecting a Playwright test execution agent with a QA agent using LangChain.
- LangChain
- Playwright
- Pydantic
- Python
This project involved fixing a complex multi-agent system. I integrated a Playwright agent (for test execution) with a QA agent (for analysis and reporting). My work focused on resolving Pydantic model conflicts and ensuring seamless communication between agents using LangChain.
- AI2025 · Upwork project
Multi-Agent QA System Config
Configured and fixed a multi-agent QA system powered by Jira and MCP, resolving Pydantic model issues to orchestrate UI/UX testing.
- Jira
- MCP
- Pydantic
- Python
For the same client, I configured a separate multi-agent system designed to orchestrate UI/UX testing by interacting with Jira. The primary challenge was debugging and resolving Pydantic model incompatibilities that prevented the agents from communicating correctly.
- Web2025 · Upwork project
AWS S3 & Django Fix
Diagnosed and fixed a critical production issue preventing a Django application from correctly connecting to an AWS S3 bucket.
- Django
- AWS S3
- IAM
- Python
A client's Django application was failing in production due to an inability to connect to its S3 bucket. I diagnosed the issue, which involved incorrect IAM policies and misconfigured Django-storages settings, and implemented a fix to restore full functionality.
- AIWeb2024 · UK client
Trading Insights Chrome Extension
A Chrome extension that analyzes trading charts and integrates with ChatGPT to provide buy/sell signals.
- Chrome Extension
- ChatGPT
- JavaScript
This was a Chrome extension that could be activated on trading websites. It captured chart data, sent it to the ChatGPT API for analysis, and then displayed the AI-generated insights (including buy/sell signals) directly on the user's page.
- Web2023 · semicolon Devs
Autocredit
Contributed to a MERN-stack loan collection tool by fixing backend logic and optimizing database queries.
- MERN
- Node.js
- MongoDB
As part of a team at Semicolon Devs, I worked on Autocredit, a loan collection tool. My role was to dive into the Node.js backend to identify and fix bugs in business logic and optimize slow-running MongoDB queries to improve application performance.
- Web2023 · semicolon Devs
SMIT LMS
Enhanced UI, fixed bugs, and added new features to an educational institute's MERN-stack Learning Management System.
- MERN
- React
- Node.js
I contributed to the SMIT (Study Medicine Italy) LMS, a platform for medical students. My work involved enhancing the React-based UI, fixing existing bugs, and implementing new features as requested by the client, which gave me practical MERN stack experience.
- Web2022 · Computer Science Society, UoP
CSUP Member Portal
The Computer Society (CSUP) member portal using the MERN stack, with user authentication, CRUD operations, and REST APIs with JWT.
- MERN
- React
- Node.js
- JWT
- MongoDB
This was one of my first full-stack projects. I built the member portal for the university Computer Society, which allowed me to learn and implement the MERN stack from scratch. Key features included user login with JWT for authentication, CRUD operations for member profiles, and a RESTful API backend.
CSUP
President
Computer Science Society (CSUP), UoP
2024–2025
Led the society: events, member collaboration, technical initiatives.
Organiser
REIGNITE ’25 Game Night
2025
Revived the annual gaming night after seven years, with an esports tournament, sponsorships and logistics.
~150 participants · 18 teams
Coordinator
“Code 2 Cloud” DevOps Workshop
2023
AWS, Kubernetes, CI/CD with Jenkins, monitoring and security.
Research
Organising Committee
RESCON, Postgraduate Institute of Science, UoP
2026
Event logistics and academic presentations.
Organising Committee
SURS, Undergraduate Research Symposium, UoP
2024
Event logistics and academic presentations.
Volunteer
Volunteer
SIRED, Society of Industrial Relations and Entrepreneurship Development
Helped run a seminar teaching mathematics to O/L students in rural schools.
Startup
Co-Founder
Xanvia
2024
AI-powered web tools and SaaS for founders. I own architecture, backend and CI/CD.
Backend architecture, AI pipelines and cloud infrastructure for startups.
Top Rated (2026)
Reviews
“Lakshitha is a very dedicated supportive intelligent developer to work with, highly recommended for projects with tight schedule. His skills in AI is excellent and delivered a very good job!”
Marianne · Upwork “He is so hardworking and good, he keeps to deadlines, Hire him if you have the opportunity you won't regret.”
Faithfulnes · Upwork “It was really good experience and Lakshitha was exceptional in his skills and very hardworking as well to explore new way around.”
Sachin · Upwork “He is the best Django guy i have worked with”
Faithfulnes A. · Upwork
Cloud Core X (Medium)
Cut Your Cloud Bills: How to Build Your Own Cloud Video Server
Mux and Vimeo bills add up fast. How we built our own HLS video streaming server for an LMS: adaptive streaming, protection against easy downloads, and full control over how videos are served.
Level Up Coding
From a 10-Minute AI Job to a 60-Second Pipeline
What happens when 50 users upload 5 videos at once? How I evolved a compute-heavy AI system from a demo-ready MVP into a production-grade pipeline without rewriting the core business logic.
Medium
Threads in C — Explained Simply with Code
Threads and multiprocessing come from the world of distributed computing. Using C, a language close to the hardware, we see how threads, processes, and memory come together.
When the laptop is closed, you'll find me with one of these.
Formula 1
Space
Motorcycles
Mountains
Beaches