I open the black box.
Compilers, coreutils, LLM calls, training loops. If I depend on it, I read it, measure it, and fix it upstream. Right now that means EnAi, a runtime optimiser that cuts a neural network's training compute while the model is still training.
01Flagship project
A training run keeps computing gradients for layers that stopped learning hours ago. EnAi watches gradient energy per layer while the run is in flight, freezes each layer the moment its contribution collapses, and raises the batch size into the headroom that frees up. All of it happens mid-run, with no restart and no change to the model code.
Solo. I build the engine, the study and the measurement.
Throughput per epoch, images per second
02Projects
Open source library2026
0.06 msadded per call
Production safeguards for LLM calls, in one decorator.
Every team putting a language model in production writes the same safeguards: retry the rate limits, cache the repeats, fail over when a provider is down, cap the spend, keep customer data out of the prompt, and make sure the answer parses. callm puts all of that behind one decorator on the function you already wrote, so the OpenAI, Anthropic or Gemini call inside it stays exactly as it was. No gateway to deploy, no database to run, and nothing required beyond the standard library.
Python · OpenAI, Anthropic and Google SDKs · SQLite · Pydantic
Case studyOpen source collective2026, ongoing
1product live
AI products, shipped free and open source.
A small group building AI products around one question: what is still harder than it needs to be? Everything ships free and open source. The first product out is Admission OPS, because university admissions is one of the most paperwork heavy moments in a person's life and most of it is deadline tracking that software should have absorbed years ago.
Next.js · TypeScript · Python · LLM orchestration
Case studyPersonal project2025
P2PWebRTC version with no media server in the path
Browser video calls, built twice: raw WebRTC and a managed SDK.
A browser video calling app in React, built two ways: a peer to peer version on WebRTC (simple-peer) with Socket.IO handling signalling and connection state, and a production version on the ZEGOCLOUD SDK. Building both made the trade off concrete: owning signalling versus shipping reliable calls fast.
React · WebRTC · simple-peer · Socket.IO
Case study03Experience
Jun 2025 to presentnow
Regensburg, Germany, Hybrid
Built the real-time voice negotiation simulator that Sales Management students train against, then re-architected it for concurrent sessions across multiple courses.
Node.js · TypeScript · OpenAI APIs · Vapi STT/TTS · Streaming audio
Sep 2025 to Aug 2026
Erlangen, Germany, Hybrid
Led the rebuild of the property websites and shipped the booking, food ordering and venue reservation flows, plus the internal operations platform the staff run on.
Next.js · React · TypeScript · Tailwind CSS · Node.js · PostgreSQL · Decap CMS
Nov 2024 to Sep 2026
Cardiff, Wales, United Kingdom, Remote
Client acquisition and the technical solutions behind delivery, plus the workflow tooling that removed the manual steps between brief and handover.
React · Next.js · Node.js · Client delivery
Sep 2022 to Apr 2023
United Kingdom, Part-time · Remote
Mobile-first web applications in JavaScript, React and Tailwind, integrating REST APIs and third-party services from wireframe through to production.
React · JavaScript (ES6+) · Tailwind CSS · REST APIs
Feb 2022 to Jul 2022
Egypt, Part-time · Remote
Frontend features in React and Tailwind with backend integration and API handling, alongside bug fixing and UI testing.
React · Tailwind CSS · Git
04Stack
EnAi, the negotiation simulator, Phoveus Lab
Every week
Now and then
TensorFlow, scikit-learn, LangChain, LangGraph, Vector databases, AI orchestration, Autograd hooks
Hotel operations platform, booking engine, REST services
Every week
Now and then
FastAPI, Express.js, Spring Boot, MongoDB, MySQL
Every product on this site, this one included
Every week
Now and then
Astro, Redux, Accessibility (WCAG 2.2), Decap CMS
Voice simulator, FaceTube. Latency is the product
Every week
Now and then
Socket.io, Speech to text, Text to speech
Roughly in order of how much I write them
Every week
Now and then
C, SQL, Rust
Getting it in front of people and keeping it there
Every week
Now and then
Postman, Figma, GitHub Actions
05Open source
06Writing
25 September 2026, 3 min read
A release of Powertools for AWS Lambda started crashing OpenAPI schema generation for any route whose response model was list[Item]. The cause was a deep copy and an identity check that could never match. Here is how I tracked it down and why the fix was one character.
23 September 2026, 6 min read
VoiceMax records your voice in the browser and tells you how you sound: the primary emotion, stress, pace, confidence and energy, plus a bit of supportive feedback. Here is how the recording, the three Genkit flows and the error handling fit together.
22 September 2026, 5 min read
FaceTube is a browser video chat with no accounts: pick a name, create or join a room, share the ID. I built the call layer two ways, once by hand with WebRTC and a Socket.IO signalling server, once on a video SDK. Here is what each one taught me.
07Background
2019Dhaka
My first program was in C, out of the ICT unit of a college syllabus. It was supposed to be one chapter, and I found it more interesting than the rest of that syllabus combined.
2021Contests
Over thirty contests and hackathons: ICPC, IUPC, university rounds, later Hackaburg and Hack | Bay in Germany. A wrong answer verdict does not explain itself, so you learn to debug your own reasoning, which is most of the job.
2023The gap
Between May 2023 and June 2024 I was not building anything. Visa applications, language and entrance exams, and a move from Bangladesh to Germany. It is the one real gap in my CV, and it seemed better to explain it here than to leave it unaccounted for.
2026Now
I did not want to put AI engineering on a CV without something behind it, so I built EnAi. The part I am proudest of is the section listing what the measurement does not prove.
An efficiency claim is only worth what its methodology survives.
Away from the screen I climb, play chess, and play guitar. None of it is a metaphor for how I work. The part that does carry over is what I want to build: software that helps people who need it should not sit behind a price, which is why everything we make at Phoveus Lab is free and open source.
08Contact
If you are hiring, want to talk about EnAi, or found a bug in one of my projects, send a message or book a call.
I read every message myself and usually reply within a day.
Book a call
Fifteen minutes, whenever suits you.
Choose any slot on the calendar. I am in Regensburg, Germany, and it will adjust to your timezone automatically.
Nothing loads from Calendly until you choose to open it.
