AI / ML Research Engineer · Malé, Maldives
Built for Dhivehi: text-to-speech, OCR, and real-time voice agents, shipped to production.
About
01 / 05A Python developer with expertise across AI systems, backend, and full-stack web. I build production AI infrastructure for low-resource languages: TTS pipelines, LLM-powered classification, and agentic workflows, deployed across FastAPI and Next.js.
Deep experience in SQL and database design, GPU optimization, and serverless deployment, shipping end to end from model fine-tuning to production APIs.
Core stack
Experience
02 / 05Work
03 / 05Text to Speech · 100k+ users
A Dhivehi text-to-speech model on a GAN architecture, trained on ~10 hours of speech. Streaming (ASGI) and standard endpoints, GPU auto-scaling, and CI/CD via GitHub Actions.
Text to Speech · Research
Fine-tuned and deployed MMS, VITS2, Orpheus and ZipVoice for Dhivehi, distilled to run on edge devices. Built Aduehy TTS, a Kotlin Android app on the Play Store that gives blind and low-vision users a Dhivehi screen reader through Google TalkBack.
OCR · DhivehiGPT
Custom detection and recognition models trained on H100 GPUs, with large-scale synthetic datasets covering three scripts at once.
Classification · Maldives Customs
Trained on 2.5M+ historical records from the Maldives Customs Service to recommend Harmonized System codes from product descriptions, served into a .NET Core app via a Flask API.
AI Agent · Open source
Watches a GitHub repo for issues labelled agent_issue and runs a coding agent end to end on each one: implement, review, fix, open a PR, merge. Every issue gets its own git worktree so many agents run at once without colliding, and a bigger diff is reviewed by a different agent or model before the PR opens. Python standard library only.
Developer tooling · Open source
An agentic skill that reads a codebase, draws C4 diagrams of it (system context, containers, components, data flows), renders them to images, then reads those images back with vision to review the architecture. The picture surfaces things that are hard to catch file by file: missing security boundaries, single points of failure, circular dependencies.
AI Agent
An autonomous agent on a Hetzner VPS with Telegram integration, multi-model LLM orchestration (Claude, MiniMax) with automatic fallback, persistent memory, and cron-scheduled content pipelines, run as a systemd service.
Infrastructure
A self-hosted serverless GPU platform on DigitalOcean Kubernetes using Beta9 and Helm, with Cloudflare Argo tunnels for secure public access.
Automation
Automated the upload and download of customs declaration documents from the ASYCUDA database, effective enough that Maldives Customs retired the Records Management Section after the backlog cleared.
Writing
04 / 05