Shamsuddin Ahmed

I’m a software engineer from Dhaka, Bangladesh, with five-plus years spent on the same stubborn question: how do you take something that works in a notebook and make it survive real users?

Most of my work sits in three places: backend systems, production machine learning, and the harness layer that lets a language model safely operate software that already exists. Lately a fourth has crept in: machine learning for biology, where the datasets are small, the labels are noisy, and being honest about your numbers matters more than beating a leaderboard.

Right now

Senior Software Engineer, Mevrik

Dhaka · since August 2026

Mevrik is an AI-powered customer-experience platform used by telecom operators and hundreds of smaller businesses. I work on the AI side of the product: agents, retrieval, and the serving infrastructure underneath them.

Founder, AlgolyzerLab

A small studio I run for applied-AI work that doesn’t fit neatly into a product roadmap: sports-science tooling for elite cricket, clinical case-management software for doctors, agriculture AI, and research engineering. Small team, real deployments, unglamorous problems.

Before that

Senior Software Engineer → Software Engineer, Evoclick · Moscow (remote) · 2023 – 2026

Deployed quantised Qwen 3.5 27B (INT4 / FP8) on vLLM for a workload serving 1,000+ concurrent users, and built the OpenAI-compatible gateway that routed traffic to it. Shipped a multi-tenant AI orchestration platform with a DAG workflow engine, layout-aware (“page-index”) RAG for document understanding, LoRA/QLoRA fine-tunes for domain tasks, and a YOLO edge-vision pipeline running on NPU hardware.

Software Engineer, OMNAIBLE · Amsterdam (remote) · 2022 – 2023

ML-powered recommendation microservices in FastAPI, plus the caching and query work that made them fast enough to keep.

Software / DevOps Engineer, CodeSmith Tech · Dhaka · 2021 – 2022

Docker-based CI/CD, hardened REST and GraphQL APIs, and IoT data-ingestion backends.

What I’m thinking about

Agent harnesses. A model is only as useful as the surface you give it. I’m convinced most teams hand an LLM the wrong thing: raw tables and auto-generated OpenAPI specs instead of the intent-named operations their own code already has. That idea drives Reins, a small harness you bolt onto an app so a model can drive it safely, and Veldra, where an agent is stored as data, a versioned spec row, rather than as code.

Machine learning for biology. I work on anti-inflammatory peptide prediction, and on structure-aware virtual screening of natural products against understudied viral proteases (DeepNatProtease). Both taught me the same lesson: the dataset you choose decides your result long before the model does.

Systems that stay explainable. Fennec (SSIM-guided image compression in Go), VoidMon (a terminal system monitor), and Clawkido (actor-model agent swarms). Small, readable, zero-magic tools I actually use.

Tools I reach for

Writing

I write to understand things, not to look clever. The posts here explain hard ideas the way I wish someone had explained them to me, with analogies, runnable code, and honest numbers. Start with the blog archive, or read the CV if you’d rather see the short version.

Say hello: info@shamspias.com · GitHub · LinkedIn