Mehemud Azad

Final-year CS undergrad at BUET · AI researcher

I research vision-language models — spatial reasoning and the mechanistic interpretability behind it — and build agentic systems for evaluation and software testing. Before research I spent two years on full-stack and DevOps: microservice platforms, CI/CD, and cloud-native infrastructure.

View work GitHub LinkedIn Email

// about

From building systems to studying them

I'm a final-year Computer Science student at Bangladesh University of Engineering and Technology. I started in full-stack web development, then spent two years on DevOps and DevSecOps — automation pipelines, cloud-native deployments, and secure delivery — alongside competitive CTFs.

My work now is research. My thesis is on spatial reasoning in vision-language models and the mechanistic interpretability behind it. I also worked on agentic systems for exploratory software testing, Bangla LLM-text detection, and hallucination and deepfake detection.

  • Vision-language models
  • Mechanistic interpretability
  • Agentic AI
  • NLP
  • DevOps & DevSecOps
  • Backend engineering

// research

Current work

Undergrad thesis — vision-language models

Spatial reasoning in VLMs and the mechanistic interpretability behind it: how these models represent spatial relations, and where in the network that computation happens.

Exploratory test automation with agentic AI

An agent that explores mobile applications on its own and generates test cases, without hand-written scripts or predefined paths.

Bangla LLM-text detection

Curating a dataset of Bangla LLM-generated text and building detection methods for it — an under-resourced setting for existing detectors.

LLM hallucination detection

Detecting hallucinated and unsupported content in language-model outputs — developed for CUET Datathon 2026.

Deepfake detection

Robust detection of AI-generated images and AI-generated video, built at BRAC AI Build Fest. The provenance-first pipeline in the Verilens extension builds on this work.

// projects

Selected projects

Verilens

Code ↗

Manifest-V3 browser extension that flags AI-generated media and misinformation on X, Instagram, and Facebook. Checks C2PA content provenance first, then falls back to deepfake models and automated fact-checking.

  • JavaScript
  • Chrome MV3
  • C2PA
  • Deepfake detection

Green Agent — GAIA Evaluator

Code ↗

An A2A-protocol evaluator for GAIA benchmark tasks, built on Google ADK. Deterministic scoring, optional LLM-based judging, and multi-agent orchestration for reliable agent evaluation.

  • Python
  • Google ADK
  • Gemini
  • A2A Protocol

Purple Agent — Build-What-I-Mean

Code ↗

A standalone LLM agent for the Build-What-I-Mean benchmark. Two-stage pipeline with speaker-aware pragmatic inference, communicating over the A2A protocol.

  • Python
  • LLM
  • A2A Protocol

CareForAll API Avengers winner

Code ↗

Champion project at CUET API Avengers 2025. An event-driven donation platform: FastAPI microservices, Kafka workflows, gRPC service communication, an observability stack, CI/CD, and Kubernetes deployment.

  • FastAPI
  • Kafka
  • gRPC
  • Kubernetes

The Freelancer

Code ↗

A microservices freelancing platform: API gateway, JWT auth, escrow payments, workspace collaboration, and vector-search matching between clients and freelancers.

  • Spring Boot
  • Next.js
  • Kafka
  • PostgreSQL

C-to-8086 Compiler

Code ↗

A full compiler pipeline written from scratch: symbol table, lexical analysis with Flex, syntax and semantic analysis with ANTLR4, and 8086 assembly code generation with optimisation.

  • C++
  • Flex
  • ANTLR4

ATmega32 Maze Solver

Code ↗

An autonomous maze-solving robot on ATmega32: left-hand-rule search, triple ultrasonic sensors, MPU6050 gyroscope feedback, and PID-assisted movement for reliable navigation on constrained hardware.

  • C++
  • ATmega32
  • PlatformIO

// milestones

Competitions & results

  1. 2026

    Kaggle AIMO 3 — Bronze medal

    Global bronze medal on Kaggle.

  2. 2026

    BUET DL Sprint 4.0 — 7th of 192 teams

    Top-4% finish in a deep-learning competition.

  3. 2025

    CUET API Avengers — Champion

    First place with Team FAT32, for the CareForAll microservices platform.

  4. 2025

    CUET Techathon — 5th place

    Hardware and IoT build under a fixed time limit.

  5. 2023

    BUET IEEE DevSprint — Rising Team

    My first inter-university competition, and the start of the competitive track.

// contact

Get in touch

The fastest way to reach me is email. I'm open to research collaboration, and to internship or new-grad roles in ML and infrastructure.