"The goal isn't to ship code.
It's to ship the right outcome."
I take complex, unclear, real-world briefs and engineer them into software that ships. Backend systems. Full-stack products. AI used deliberately. Every line owned.
Client requirements, vague briefs, stakeholder notes. I map the domain and the real constraints before a single schema gets designed. Most "technical" problems are scope problems in disguise.
→ domain mapping firstThe data model is the architecture. I design all entities and relationships before any interface decision. The database shape determines everything downstream.
→ structure before interfaceV1 is not the vision. It's the minimum that proves the concept with real data, real auth, real workflows. Everything else is sequenced — not dropped.
→ scope ruthlesslyAuth → data layer → API contracts → frontend consuming real endpoints. No mocked data in production code. No UI before the backend is real.
→ foundation before featureDeployed software teaches you things no prototype can. The first deployment is where real feedback starts — not the finish line.
→ ship then iterate
"The goal isn't to ship code.
It's to ship the right outcome."
AI amplifies both good thinking and bad thinking. In a messy project, it generates confident-sounding nonsense if you let it run without constraints.
My workflow uses AI as a thinking partner under strict rules — not as an output machine. Every decision that ships is mine to own.
I describe the messy brief in full before asking AI anything. Vague input produces confident garbage. Precision in the prompt is precision in the output.
I ask for three different approaches to a problem, then evaluate trade-offs myself. AI gives breadth. I provide judgment. The decision is always mine.
I never let AI design the data model. The database schema is the most load-bearing decision in any project. I own it completely. AI doesn't touch it.
When I'm stuck, I describe the bug to AI in full. Articulating the problem clearly usually reveals it. AI is a thinking mirror — not a fix machine.
I get the feature working with my own logic first. Then I ask AI to review for patterns and readability. Never the reverse — functionality is not negotiable.
If AI suggests a change, I read it line by line and understand it before it enters the codebase. I don't ship code I can't explain.
Real projects, real problems. Click any build to see the chaos that went in — and what came out.
Every AI task gets sent to the same expensive model, even trivial ones. Teams overpay for intelligence they never needed — and nobody had built a router that actually proved its own routing decisions were correct.
A 4-layer complexity classifier routing tasks to a free local model (Ollama/ROCm) or a remote model (Gemini/Fireworks) with confidence-based escalation, live token streaming, a real-time dashboard, and a self-graded 50-task benchmark suite producing a reproducible VoxRouter Score.
AI agents can now propose and push real changes — code, PRs, Slack messages — with no gap between "the agent decided this" and "it's live." No staging, no cryptographic proof of what's about to happen, no safe way back out.
A local-first transaction layer: agent-proposed changes get staged in isolation, verified deterministically, and human approval is cryptographically bound to the exact digest of that change — only then does it release through controlled adapters (GitHub branch, draft PR, Slack), with idempotent receipts and crash-safe recovery. 70 Go binaries, own CLI, cooperative and rootless-OCI execution modes.
A real delivery platform needs more than parcel forms. It needs role-based flows, secure payments, rider assignment, admin control, and deployment discipline across frontend and backend.
Deployed MERN parcel delivery system with Firebase authentication, user/admin/rider workflows, Stripe payment flow, parcel status handling, rider approval, and admin parcel management.
Built role-based access around real user workflows
Production deployment flow with Firebase hosting
Manual attendance fails in ways that are hard to audit. Face recognition must work across lighting variance, angles, and a growing face database — reliably, in constrained conditions.
FaceNet (InceptionResnetV1) + MTCNN detection, cosine similarity matching, Flask REST API decoupling the ML inference from the frontend, vanilla JS dashboard for real-time attendance logging.
Looking for engineering roles and projects where requirements are messy, the stakes are real, and "it works" actually means something. Based in Dhaka. Open to remote.
Start a conversation →