About
From Sri Lanka to Japan, building AI that ships.
I'm Vihanga Nimesha, a software engineer who spends most of his time building AI agents and LLM systems in Python: custom agent loops, retrieval (RAG), fine-tuning, and evaluation. I also work full-stack with Next.js and FastAPI, and on cloud and edge infrastructure (AWS, GCP, NVIDIA Jetson) when getting a system into production calls for it.
A short bio
My path into engineering started in Sri Lanka, with a stubborn curiosity about how the systems around me actually worked underneath the surface. That curiosity turned into a degree: a BSc (Hons) Software Engineering awarded by the University of Plymouth, which I studied at NSBM Green University between 2020 and 2023.
From there I was selected for a competitive METI internship, a Japanese-government program, which became a permanent role. Moving countries was its own kind of engineering problem: a new language, a new culture, a whole set of unwritten rules. It was hard, and it was the best thing that could have happened to my craft. It made me more deliberate, more patient, and more ambitious about the work I want to do.
Today I'm a software engineer at Brycen in Japan, where my work moves between web apps, applied AI, and edge devices. A good example is an internal knowledge-base chatbot I built: a Next.js front end over a Python (FastAPI) service that lets staff ask questions of company documents, answered with retrieval-augmented generation against a Milvus vector store. I've also prototyped agents and fine-tuned LLMs with Hugging Face and TRL, and taken YOLO object-detection models onto NVIDIA Jetson boards, where converting them with TensorRT brought a single inference down to roughly 10.5 ms (about 95 FPS), some six times quicker than running on CPU. To keep all of this reproducible, I provision our AWS environments with Terraform, so a merge to main brings a fresh stack up in minutes.
Building AI agents is the work I'm drawn to most, so I keep at it outside work too. I built a multi-agent LLM platform from scratch: a custom async streaming agent loop with sandboxed tools and a vector-backed memory layer, and got obsessive about keeping it efficient, cutting token usage per conversation by more than half (around 110k down to 50k) through prompt and context optimisation. That kind of work has given me a practical sense of where today's models genuinely help and where they fall short.
What I enjoy most is owning the whole arc of a system: the model, the infrastructure it runs on, and the interface people actually use, while tying my technical choices back to a measurable outcome: cost, latency, or hours saved, not just features shipped. To me, technology has always felt like an open frontier, where every answer opens another door and curiosity keeps doing the rest.
Things I've built
Multi-Agent Chat Platform
End-to-end build
A multi-agent LLM chat platform with a custom streaming agent loop written from scratch: per-tool timeouts, retries, backoff, and output truncation. Sandboxed tool layer (Bash, Read, Write, Edit) where the model writes and runs its own Python to produce real, downloadable XLSX, PPTX, PDF, DOCX, and Markdown files. SSE streams thinking, text, and tool calls into the UI as interactive cards, with hybrid model support across OpenAI and Ollama's Anthropic-compatible endpoint, context-aware memory via Milvus, and an n8n webhook layer.
Next.js 16React 19TypeScriptFastAPISupabaseMilvusFederated Learning Framework for LLMs
Research project
A custom FedAvg-based federated learning pipeline for large language models, training multiple LLMs independently and aggregating weights into a unified model, enabling privacy-preserving collaborative training across decentralized environments.
PythonPyTorchHF TransformersSpiking Neural Networks
Research project
A 3-layer LIF spiking network in snnTorch/PyTorch using rate coding, surrogate gradients, and 25 timesteps, reaching 98% on MNIST. Analyzed spike dynamics and membrane integration to understand event-driven inference for neuromorphic hardware.
snnTorchPyTorch
Featured Posts
Building something with AI? Let's talk.
I'm always interested in conversations about LLMs, agents, computer vision, and end-to-end AI systems, whether you're shipping a product, doing research, or just exploring.