AI Engineer / Agentic Systems
Building systems that plan, retrieve, and act — not just predict.
I design and ship agentic AI: multi-step reasoning pipelines, retrieval-augmented research tools, and the orchestration layers that hold them together.
4 agent roles · self-verifying retrieval · Shahabuddin's research assistant
01 / About
Why agentic systems, and not just models
I got interested in AI for a boring reason: most software I used just executed instructions, and I wanted to build something that could decide what to do next. The first time I got a basic retrieval pipeline to reason over its own search results instead of just returning them, I was hooked on the gap between 'AI that answers' and 'AI that works'.
That gap is where I spend most of my time now. At OrangeBD I work inside that gap daily — chatbots that need memory, recommendation systems that need judgment, agents that need to verify their own output before handing it back. None of it is glamorous in the way demos are. It's mostly state management, prompt failure modes, and re-reading LangGraph traces at 1am to figure out why an agent looped twice.
I think the next decade of software is going to be defined by systems that can plan and act with some autonomy, and I want to be one of the engineers who actually knows how to build the unglamorous parts of that reliably — not just prompt a demo, but ship something that holds up.
02 / Experience
Where the theory meets production
Jan 2026 — Present
Trainee Software Engineer (AI)
OrangeBD
Working inside the AI team on production ML/AI systems — chatbots, retrieval pipelines, and recommendation logic — while collaborating cross-functionally to ship intelligent features end to end.
03 / Projects
Systems I've shipped, not just demoed
Retrieval that plans before it answers
Agentic RAG Research Assistant
A research assistant that treats retrieval as a multi-step decision process rather than a single lookup. Autonomous agents plan a research strategy, retrieve from multiple sources, cross-verify claims against each other, and synthesize a final answer — instead of returning the first plausible match.
Pipeline stages
4
Agent roles
Plan · Retrieve · Verify · Synthesize
Challenge
Naive RAG fails silently when retrieved context is wrong or incomplete — it answers confidently anyway. The hard part was building a verification step that could catch contradictions between sources before synthesis.
Lesson learned
Most of the reliability gain came from giving the agent the ability to say "this evidence is insufficient" rather than from a smarter prompt.
One chatbot, several jobs, one source of truth
Multi-Agent AI Assistant
A multi-agent RAG chatbot handling order management and employee leave workflows behind a single conversational interface, backed by JWT auth, an admin dashboard, and a lightweight embeddable widget.
Workflows handled
2
Core services
Postgres · Redis · OpenSearch
Challenge
Routing a single conversation across multiple agents (orders vs. leave requests) without the user noticing a handoff happened, while keeping auth and state consistent across both.
Lesson learned
Session state needs to live outside any one agent — Redis as a shared scratchpad mattered more than any individual agent’s logic.
Role-based academic operations, built in Java
University Management System
A complete university portal built with an object-oriented Java architecture, with role-based dashboards for administrators, teachers, and students covering academic and administrative operations.
User roles
3
Dashboards
Admin · Teacher · Student
Challenge
Modeling shared academic data (courses, grades, schedules) so three very different roles could see the right slice of it without duplicating logic across dashboards.
Lesson learned
This was the project that taught me the value of designing the data model before the UI — it kept paying off every time a new role-specific view was added.
04 / Stack
The tools the systems are built from
AI / ML
90%Backend
80%Databases
75%Languages
85%Tools
78%05 / Leadership
Leadership beyond engineering
Coordinating people has turned out to matter as much as coordinating agents.
7th
National Science Carnival edition led
President
Dhaka College Science Club
2021 — 2022
Led organization of the 7th DCSC National Science Carnival — the country's largest national science event — supervising an inter-college science exposition that promoted innovation and scientific curiosity.
2
Flagship events organized
Associate Executive
North South University HR Club
2024 — Ongoing
Played a key role organizing Calibration 4.0 and Epitome 3.0, contributing to event planning, team coordination, and partnership communications.
Active
Social development campaigns
Institute Representative
Volunteer for Bangladesh
Ongoing
Collaborated with teams to organize awareness campaigns and contribute to sustainable social development projects.
06 / Activity
Live from GitHub
—
Repositories
—
Followers
—
Following
07 / Contact
Let's build the future together
Open to roles and collaborations in agentic AI, applied ML, and systems that need to reason — not just respond.