I am a Senior Full Stack Software Engineer in Badda, Dhaka, Bangladesh. Most of what I do sits underneath the product - data models, work queues, delivery guarantees, tenancy boundaries and the pipelines that carry everything between them. I also ship the UI and the deployment, because the person who designed the contract should be the person who finds out it was wrong.
Four years ago that meant Django views. Today it means deciding whether a lease should be five minutes or fifteen, and knowing what breaks either way.
Core stack: Python, Django, FastAPI, PostgreSQL, Redis Streams, Celery, pgvector, Docker, Traefik, AWS, Next.js, React Native, LangGraph, Scrapy, Playwright, Prometheus, Grafana, Cloudflare, PgBouncer, WebSocket.
Each of these cost me something to learn. They show up in every system on this site.
The happy path is a weekend. What takes real engineering is deciding what happens when a worker dies mid-task, a provider rate-limits, or the same message arrives twice. I start there now.
Anything that depends on a developer remembering a filter will eventually be forgotten. Schema-per-tenant, queryset-level scoping, content-derived idempotency keys - put correctness where it cannot be skipped.
Cost-order your pipeline. Cheap checks absorb the volume; the expensive method only sees the residue that genuinely needs it. This applies to deduplication, authorization and LLM calls alike.
Link duplicates instead of deleting them. Deploy the previous tag instead of hot-fixing. Every threshold I have ever picked was wrong the first time, so I make being wrong cheap.
LLMs are excellent at intent and unqualified for authority. Schema-validated proposals, permission checks the agent cannot bypass, and a blast-radius ceiling - that boundary is what makes AI features shippable.
Architecture docs, decision records and mentoring are not overhead. They are the difference between a system one person can run and a system a team can own.
Four years that ended in a thesis on Bengali-to-English word alignment using transformer-based multilingual BERT - my first real encounter with the gap between a model that works in a notebook and one that works.
Built the incubator's own platform while advising the startups inside it, and began an MSc in IT at IIT, University of Dhaka - which I have been carrying alongside full-time engineering ever since.
Went from executing tickets to shaping systems in under a year: requirements with leadership, architecture decisions, and end-to-end ownership across several products. Started leading juniors.
Communication infrastructure across email, push, SMS and WhatsApp; a multi-branch attendance engine with payroll integration; and integrations with SSLCommerz, Stripe, Twilio and Google OAuth.
Technical leadership, monolith-to-microservices decomposition, multi-tenant CI/CD provisioning, and AI-driven workflows in the internal development loop. Master's thesis shipping as a production platform.
Recruiters search for titles; systems do not have them. Here is how my work maps onto the labels you are probably searching for.
One person, the whole vertical slice.
I take a feature from an ambiguous requirement to a deployed, observable slice - schema, API, UI state, background jobs and the release that carries it. Three products in production, two of them multi-tenant SaaS with a web app and a mobile client on the same API.
Correctness under concurrency is the whole job.
Most of my hardest work lives below the API: lease-based work queues, at-least-once stream processing, schema-per-tenant isolation, and the transactional edges where those meet. I reach for Postgres primitives before I reach for another service.
LLMs are unreliable subsystems. I engineer around that.
My NLP background is academic - a Bengali↔English word alignment thesis on multilingual BERT - but my LLM work is operational: schema-validated extraction, provider fallback, cost tracking, versioned embeddings and semantic search that has to answer in production latency.
If I built it, I can deploy it, watch it and roll it back.
I own the path from a merged commit to a running container behind Traefik and Cloudflare - multi-tenant provisioning included. Zero-downtime is a deployment property I design for, not a thing I hope for.
The part that outlives the sprint.
Reusable abstractions, written-down architecture, tests where they earn their keep, and mentoring so decisions survive me. A Master's in IT alongside full-time delivery, and an open-source package on PyPI that forced me to take versioning seriously.
IIT, University of Dhaka
Thesis (Project): Agentic Email Campaign Management Platform
North South University
Thesis (NLP Research): Bengali to English Word Alignment using Transformer based Multilingual-BERT
I read everything that comes in. If you are debugging something in this territory and want a second opinion, that counts as a reason to write.