Photo of Md Shihab Hossain Open to full-time and contract work · Remote

I build AI software and make it work for real users.

Hi, I'm Md Shihab Hossain, a software engineer. I build AI and full-stack products end to end, and I fix what breaks close to the people using them.

Focus
AI products · Full-stack
Stack
Python · TypeScript · Next.js
Data
Postgres · Supabase · pgvector
Open to
Full-time · Contract · Remote
Based
Dhaka · UTC+6
Hours
Overlap with US mornings
01 · About

An engineer who sits with the problem

I'm an AI engineer who does my best work next to the people using the product. I sit with the problem, learn how their work actually runs, and build what they need, whether that's an AI feature, an integration or a fix deep in a codebase I didn't write. I came up through competitive programming, which is where I learned to break a hard problem down and get it right under pressure. Since then I've built AI products end to end, shipped fixes into open-source projects, and started a few companies of my own, including a small dev studio and the GTM automations that found its clients. I'm happiest when I can talk to a user in the morning and ship their fix by the evening.

02 · Open source

Fixing what breaks for real users

I contribute fixes to Agent Duel, an open-source arena where two AI coding agents take the same task and users vote on the result. I made long battle summaries keep each agent’s final answer, reopened chats saved before the switch to Arena-only models, removed a phantom empty row from the Changes view, and fixed the sandbox’s write check twice, so rm -rf no longer reads as a write to a file named f and a path after -- can’t slip past it.

03 · Projects

Things I've built and shipped

CareLoop · family dashboard
The CareLoop dashboard: today's medication schedule and a scanned pill-bottle label waiting for confirmation
Computer vision + AI · Private

CareLoop 98% of medicines read from real prescriptions

Snap a pill bottle. Get the day's medication plan, with reminders by voice.

An AI medication companion for families caring for aging parents. CareLoop reads a pill bottle or prescription label from a photo, checks it against the current medication list and builds the daily schedule. Alexa+ keeps the parent on track at home, and the family hears about missed doses.

  1. Photo of a label
  2. OpenCV flattens it
  3. Vision model reads it
  4. Reasoning model checks it
  5. Schedule, voice, alerts
  • Tested on real prescriptions. On handwritten Bangla and English prescriptions shared by families, it found nearly every medicine. Its mistakes drove the next fixes, so it now reads only the latest visit and only the reading step can ever set a dose.
  • Reads messy photos. Flat labels get perspective correction, and curved pill bottles get a cylindrical unwrap.
  • Never medical advice. Duplicates, the same ingredient under two brand names and timing mismatches go to the pharmacist. Doses are never changed.
  • Works through Alexa+. An MCP server answers "What do I take next?" and logs doses, and the family gets one alert per missed dose.
Vision
OpenCV · vision model
Reasoning
Nemotron on Nebius
API
FastAPI
Voice
Alexa+ via MCP
Store
SQLite
Deploy
Docker
Cogniflow
Cogniflow answering a question about a YouTube video, with each claim cited to the second
RAG · Open source

Cogniflow AI 27 stars · 13 forks on GitHub

Ask your videos, podcasts and PDFs. Get cited answers.

Retrieval-augmented generation that runs in your browser. Add YouTube links, podcasts, PDFs or notes; every claim in the answer carries a citation that links back to the source, and for YouTube to the exact second.

  1. Videos, podcasts, PDFs
  2. Chunk + embed in the browser
  3. Top-k search
  4. Gemini answers
  5. Every claim cited
  • Local-first. Sources, chunks and embeddings live in IndexedDB; only small excerpts go to the model.
  • Retrieval that holds up. Paragraph-aware chunking with overlap, task-typed embeddings and top-k cosine search.
  • Honest refusal. If the sources don't contain the answer, it says so instead of inventing one.
LLM
Gemini 2.5 Flash
Embeddings
gemini-embedding-001
Retrieval
Top-k cosine
Chunking
Paragraph + overlap
Store
IndexedDB
App
Next.js 15 · TypeScript
DevChat
DevChat channel with a shared code snippet and an AI explanation under it
Real-time + AI · Open source DevChat CI status

DevChat

Team chat for developers, with AI code explanations in the thread.

Share code in a channel, get it highlighted properly, and click Explain to stream an AI breakdown under the snippet. There's no custom backend. The browser talks to Supabase directly, and Postgres row-level security is the authorization layer.

  1. Message or code
  2. Postgres checks access (RLS)
  3. Realtime to teammates
  4. Explain streams from OpenAI
  • Security tested, not assumed. Every allowed and forbidden action is tested against the real database's RLS policies.
  • Realtime without a socket server. Postgres change feeds handle messages, presence and typing, with backfill after reconnects.
  • Bring your own key, safely. OpenAI keys are encrypted with AES-256-GCM and never sent back to the browser.
  • One-click guest demo. Guests sign in anonymously, and a database cron job cleans up after 24 hours.
App
Next.js 15 · TypeScript
Realtime
Supabase Realtime
Auth
Postgres RLS · Google
AI
OpenAI, streamed (SSE)
Code
Shiki · Monaco
CI / Deploy
GitHub Actions · Vercel
Outbound Engine · review queue
The Outbound Engine review queue: drafted first lines waiting for approval before anything reaches the CRM
GTM automation · Private

Outbound Engine

Research and first drafts, done before you sit down.

A webhook- and cron-driven prospecting pipeline. It pulls prospect and company data via Apify, summarises recent company news with OpenAI, drafts a personalised first line for each person, and pushes qualified leads to the CRM.

  1. Webhook or cron
  2. Apify sourcing
  3. News summary + first line
  4. Pydantic checks
  5. CRM, for review
  • End to end. Sourcing, research, drafting and CRM sync run in one pipeline.
  • Validated at every stage with Pydantic, and a PostgreSQL log of every run.
  • Human stays in control. Drafts land for review, and nothing is sent automatically.
Data
Apify
AI
OpenAI
Schema
Pydantic v2
Trigger
Webhooks + cron
Store
PostgreSQL
CRM sync
REST API

I've also built a brand-mention monitor, a support-ticket triage bot, a voice booking agent, an invoice extraction pipeline and a PRD-to-tasks tool. Happy to walk through any of them.

04 · Experience

Where I've worked

  • Aug – Oct 2026

    Backend Software Engineer Intern

    Springer CapitalRemote, US

    Node.js and Python services and REST APIs for real estate valuation and cash flow modeling, on PostgreSQL with Redis task queues.

  • 2024 – 2026

    Founder

    NextOrbit LabsDhaka

    Small dev studio building websites, apps and MVPs. I ran both sides, finding the work and shipping it.

  • 2024

    AI Data Trainer

    OutlierRemote

    RLHF evaluation of LLM output across coding and reasoning tasks.

05 · Education

Education

  • Bachelor of Science in Computer Science and Engineering

    North South UniversityDhaka
06 · How I work

Close to the user, careful with the code

  • Solve the hard part first

    Competitive programming taught me to find the core of a problem, prove the idea, and only then write the code.

  • Start from the user's problem

    I use the product, hit the problem myself, and reproduce it before I touch the code.

  • Find the root cause

    I find my way around large codebases I didn't write, and fix the cause, not the symptom.

  • Ship small, tested changes

    Each fix is as small as it can be and comes with a test that proves it.

  • Write it down clearly

    Bug reports, pull request descriptions and updates a busy team can act on without a meeting.

  • Use coding agents with judgment

    I've spent months judging AI-written code, at Outlier and in open source, so I know where agents help and where to check their work.

Because I've run a company myself, I care whether a fix matters to the business, not just whether the ticket is closed.

07 · Skills

Skills

Languages

  • Python
  • TypeScript
  • JavaScript
  • SQL

Frontend

  • React
  • Next.js (App Router)

Backend

  • Node.js
  • FastAPI
  • Pydantic
  • REST APIs
  • Webhooks + cron
  • Streaming (SSE)
  • Row-level security

Databases

  • PostgreSQL
  • Supabase
  • Redis
  • SQLite
  • pgvector

AI & LLM systems

  • OpenAI API
  • Claude API
  • Gemini API
  • MCP
  • RAG
  • Function calling
  • Structured outputs
  • Embeddings
  • OpenCV

Infrastructure

  • Docker
  • Vercel
  • GitHub Actions

Automation

  • n8n
  • Apify
  • Puppeteer
  • CRM APIs

Developer tools

  • Git
  • GitHub
  • Claude Code
08 · Writing

Writing

Notes on what I build and what I learn shipping it.

09 · Contact

Let's talk.

Need an engineer who works close to your users? Email me or book a short call. I'm in Dhaka (UTC+6), so my evening overlaps US East mornings and UK afternoons.