Seeking PhD opportunities3D vision, navigation & language models
PhD applicant · AI/ML Engineer · IEEE-published researcher
I build computer-vision and LLM systems that show their reasoning, then ship them.
About
I’m a computer science student at BRAC University who builds AI meant to work outside the lab. Most of my work sits where computer vision meets people who need it to be dependable.
That includes a first-author IEEE Access paper that recognises signs of stroke from facial landmarks and explains each prediction, and an assistive vision agent for blind and low-vision users that decides when to answer, hedge, or ask for a better photo.
On the Mongol Tori Mars Rover team I designed PCBs and tuned onboard detection. Lately I build LLM systems too, like a document assistant that cites the page behind every answer.
What connects it all: AI that shows its reasoning and admits what it doesn’t know.
Research & skills
Four research directions, each with the papers and systems behind it. The toolkit underneath is what that work runs on.
Metric, physically consistent 3D from ordinary cameras, and policies that use it to move through the world.
Calibrated confidence and explanations that stay stable, applied where mistakes cost the most: health and safety.
How transformers hold long context and meaning: positional encodings, sentence embeddings, and retrieval that cites its source.
Perception and agents for people and machines acting in the real world, with fixed safety rules when the evidence is weak.
Experience
Current – Present
BRAC University · Research Seed Grant Initiative 2025 · Dhaka
–
BRAC University · Computer Science & Engineering · Dhaka
–
BRACU Mongol Tori · Mars Rover Team · Dhaka
–
BRACU Dichari · Rescue Rover Team · Dhaka
Selected work
Research prototypes and shipped systems. The full index has 22, searchable and grouped by research area.
Ongoing researchFeatured project
An agentic vision assistant for blind and low-vision users. On-device tools answer most questions; a calibrated-competence policy decides when to answer, hedge, ask for a better photo, or decline. A cloud vision-language model is called only when the local tools can’t settle it.
Answers HR policy questions with the document, page and quote behind each answer, or says it doesn’t know. It also checks company policy against labour law and flags the gaps.
YOLOv8 finds the keyboard and PaddleOCR reads the keys so the rover’s arm can type on its own. Part of the vision work that improved on-rover inference by up to 30%.
Touchless text entry: MediaPipe hand landmarks drive a virtual keyboard and mouse, with bilingual English/Bangla next-word prediction. 97.4%+ gesture accuracy at up to 37.7 FPS.
Fuses SegFormer road segmentation, YOLO detection and MediaPipe pose to predict conflict risk in Bangladeshi road scenes. XGBoost reached 95.48% accuracy on RSUD20K, explained with SHAP.
ArUco-marker navigation to ±5 cm and a vision-guided 2-DOF arm for pick-and-place, split across two ESP32s so each subsystem fails on its own.
A MERN platform with role-based dashboards, appointment booking with video consultations, and electronic medical records with e-prescriptions.
Research
4 peer-reviewed papers across explainable AI, physiological signals and touchless interaction, cited 38 times.
A real-time touchless keyboard-mouse pairing MediaPipe hand landmarks with two-layer LSTM next-word prediction in English and Bangla.
Recognises stroke from facial landmarks with MediaPipe and machine learning, with SHAP explanations so clinicians can see why each prediction was made.
Classifies four sleep stages from low-cost PPG signals with explainable ML, making interpretable sleep monitoring viable on wearable-grade sensors.
Benchmarks ML algorithms for PPG-based sleep staging and maps the accuracy and efficiency trade-offs for signal-only setups.
Recognition
Competitions, academic honours and certifications, newest first.
2025 · Competition
University Rover Challenge 2025 · The Mars Society, with BRACU Mongol Tori
Certificate of AppreciationCSE471 Project Showcase, Fall 2025 · Department of Computer Science & Engineering, BRAC University
First Runner-UpSystem Analysis & Design Project, Summer 2025 · School of Data & Sciences, BRAC University
Third Runner-Up & Grand FinalistEXCELERATE 2025 · Finance & Accounting Club, BRAC University · team SlayerX
Second Runner-Up, RobosoccerTECHSPECTRA · Robotics Club of BRAC University
Bronze AwardThe Duke of Edinburgh’s International Award · May 2023 – Mar 2024
Semi-FinalistHult Prize BRAC University 2023 · On-campus rounds
Beyond engineering
–
Notre Dame Cultural Club (NDCC)
Led 30 executives and 70 volunteers, grew the audience tenfold with 20+ partner institutions, and beat fundraising goals by 40%. Started Virtual Cultural Showdown 2020, Bangladesh’s first virtual cultural contest.
–
BRAC University Cultural Club (BuCuC)
Ran marketing campaigns, digital platforms, and the archive and photography for the club’s events.
Contact
Open to AI/ML engineering roles, research positions and collaborations across computer vision, agentic systems and MLOps.
