Munim Thahmid

Munim Thahmid

Final-year CS undergraduate at BUET. Working on reliable and secure software systems, AI for software engineering, and LLMs for code.

Email GitHub Scholar CV LinkedIn
About

I am a final-year Computer Science & Engineering student at Bangladesh University of Engineering and Technology (BUET). My research interests span reliable and secure software systems, AI for software engineering, and LLMs for code, with a particular focus on software reliability and testing.

My thesis, advised by Dr. Sadia Sharmin, interprets multilingual speech models for Bengali through layer-wise probing across Whisper and XLS-R. Alongside academic work, I spent over a year at Yobo AI building backend systems and testing infrastructure for production AI voice-agent applications.

Reliable & Secure Software Systems AI for Software Engineering LLMs for Code Software Reliability & Testing
Publications
Interpreting Multilingual Speech Models for Bengali: Cross-Model Probing and Representation-Guided Adaptation
Munim Thahmid, Sadia Sharmin
Manuscript in preparation
Selected Projects
Research
ConfigDiffGuard
A Python CLI that compares old and new vLLM configurations and flags risky changes before deployment. Curated 36 configuration-sensitive vLLM GitHub issues into three reliability-oriented failure families, and implemented YAML/JSON parsing, tracked-parameter diffing, 12 warning rules, and scored risk reports — with 5/5 intended evaluation cases passing.
Python vLLM Static Analysis
Bengali Speech Representation Analysis & ASR Adaptation
BUET thesis on interpreting Bengali representations in multilingual speech models using FLEURS, Common Voice Bengali, and MMS forced alignment. Layer-wise probing across Whisper and XLS-R showed intermediate layers strongest for Whisper language separation, and XLS-R-300M substantially stronger than tested Whisper variants on Bengali aligned-unit probing, reaching macro-F1 0.6884 in the pilot setup.
PyTorch Whisper XLS-R FLEURS MMS
Industry
Loopz.ai ↗
Production voice AI platform with real-time conversational agents. Architected the end-to-end voice pipeline, LLM integration, and telephony infrastructure.
Vapi FastAPI React Supabase ElevenLabs
AssistMind.ai ↗
AI-powered voice assistant platform with configurable agents, Stripe billing, and real-time analytics dashboard. Built full backend and billing infrastructure.
Next.js Supabase Stripe LiveKit Deepgram
HelloYobo Phoneline ↗
Voice AI phone system for automated customer interactions. Designed the call routing logic, speech pipeline, and conversation state management.
Vapi FastAPI Deepgram
Experience
Software Engineer Intern → Software Engineer (Part-time) Oct 2024 — Feb 2026
Yobo AI — Remote
Built backend systems, testing infrastructure, and production features for AI voice-agent applications. Developed an automated Python testing platform that reduced regression cycles from 8 hours to 15 minutes, enabling faster and more reliable deployment.
Undergraduate Thesis Researcher 2025 — present
BUET — Dept. of Computer Science & Engineering
Thesis on interpreting multilingual speech models for Bengali, advised by Dr. Sadia Sharmin. Cross-model probing across Whisper and XLS-R, with representation-guided adaptation for low-resource ASR.
Education
B.Sc. in Computer Science & Engineering
Bangladesh University of Engineering and Technology (BUET)
Jan 2022 — May 2026