01Elorva
The online store for a luxury fragrance retailer in Bangladesh, with the stock room and the order automations behind it. More than BDT 3 million in revenue has gone through it since launch, and my team still runs it.

I build software people run their work on: an online store past BDT 3 million in sales, my own agency platform with AI agents, and an open source desktop app. Founder of Octarnal, still a CSE student.Muntahi SafwanView CV ↗
I take a product from an idea to something people use every day, with a small team in Dhaka . Stores that sell, platforms a business runs on, AI that shows its work, and tools that make a computer feel lighter, not heavier.
Products that are live and in use. Elorva has taken more than BDT 3 million in revenue. Octahive is open for sign-ups.
A lossless audio codec paper in revision at a Springer Nature journal, and work on making GraphRAG cheaper.
Cloister is open source, and I write down what the work teaches me.
The online store for a luxury fragrance retailer in Bangladesh, with the stock room and the order automations behind it. More than BDT 3 million in revenue has gone through it since launch, and my team still runs it.
A directory SaaS for a clinic group in Australia. Public pages where people look up a therapist near them, and an admin dashboard for managing every listing.
Quotes, projects, invoicing, books and a client portal in one workspace. AI agents draft the routine work, and a person approves anything before it leaves.
WhatsApp, Discord, Gmail, Messenger and any website in one lightweight window. Written in Rust with Tauri, private by default, with a Windows installer of about 5 MB.








Things I explore when there is no client brief. Mostly about making systems smaller, faster and more honest about what they know.
Using reinforcement learning to shrink a knowledge graph to a token budget while keeping the parts an answer actually depends on.
Research · 2026L/02An optimized block Huffman scheme for real-time lossless audio, written up as a journal paper with my co-authors.
Paper · 2026An AI assistant that answers from a company's knowledge graph reads far more of it than any one question needs, and every token it reads costs time and money. This toy shows the idea I am researching: shrink the graph, keep the answer.
A list of logos says very little. This map links each tool to the work it was actually used in. Hover a node, drag it around, or filter by area.Tap a project to see what it was built with, or a tool to see where it was used.
Harvard's introductory computer science course, completed online before university.
Began my degree and took on paid work with a friend who is now my co-founder. Learned the job by doing it.
A year of full-stack work on a team, from September 2024. Next.js, React, Node.js and PostgreSQL.
Named to the Dean's List at AIUB five times so far, while studying on an academic scholarship.
On the research and development team, building machine learning systems that run on hardware, alongside the studio.
The work became a software agency. A small team in Dhaka doing design, web, mobile and AI systems.
Orders, production and sales for a bakery in Bangladesh. It taught me more than any course.
Platforms shipped for clients in the USA, Europe and Australia.
A journal paper with my co-authors, submitted to a Springer Nature journal and currently in revision.
Co-authored research using reinforcement learning to make retrieval cheaper without losing the answer.
Octahive is where the two meet: a product people pay for, with AI agents doing real work inside it.

I run Octarnal alongside my CSE degree at AIUB. It is a small team, and most of what I know about shipping software I learned from the people in it: how to scope honestly, how to say no to the wrong feature, and how to keep a client in the loop when things slip.
I like problems where the product and the engineering have to be solved together. Multi-tenant platforms, AI that has to be right, apps that need to feel fast on ordinary hardware. I am not finished learning any of it, and I would not want to be.
An AI agent that knows my work and can drive this page for you. It answers only from a fact sheet I wrote, and tells you when it doesn't know.
Answers are generated by AI from a fixed fact sheet. For anything that matters, ask me directly.