hi there!
I'm Abdul Basit Tonmoy,
an engineer, researcher, hacker & photographer a coffee & sunset enthusiast a weekend hiker & amateur footballer
I think about how learned representations deform under fine-tuning, quantization, and distribution shift — and how to keep them cheap at scale. Currently co-founding Skop.
email github linkedin cv resume
currently brewing
- Building Skop.
- AdaFrame is under review at ACM Multimedia 2026.
- SemanticSentry is under review at BlackboxNLP @ EMNLP 2026.
- reading on representation geometry, post-quantization drift, and physics-informed ML.
02 : Publications see all →
-
Geometric Drift Metrics are Insufficient: A Matched-Magnitude Dissociation Between Aligned and Anti-Aligned Fine-Tuning
-
AdaFrame: Hierarchical Multimodal Deduplication with Adaptive Information Budgeting for Cost-Efficient Video Advertisement Analysis
-
Advanced Facial Emotion Classification with 135 Classes for Enhanced Cybersecurity Applications
03 : Selected Projects see all →
-
SemanticSentry
2025 · libraryOpen-source library accompanying our BlackboxNLP submission. A model-agnostic framework for detecting semantic drift in embedding spaces — quantifying geometric transformations induced by fine-tuning, quantization, and LoRA updates without requiring labeled evaluation data. Combines CKA for global structure, NPS for local neighborhoods, and Isotropy for spectral geometry, accelerated by FAISS nearest-neighbor search and surfaced through W&B and MLflow hooks.
- python
- pytorch
- faiss
- transformers
- clip
- w&b
- mlflow
-
AdLovin
Oct 2025 · CalHacks-12 winnerA multi-agent advertisement analysis pipeline combining computer vision, vector retrieval, and LLM reasoning to extract marketing and business signals from video. Reaches 70–90% frame deduplication and reduces vision-API cost by roughly 70% while preserving downstream extraction quality.
- python
- claude api
- chromadb
- opencv
- pytorch
- hubert
- ffmpeg
04 : Experience see all →
-
Machine Learning Researcher
Jan — Apr 2026Wabash College
Built AdaFrame, a three-tier deduplication cascade (perceptual hash → LPIPS → CLIP clustering) with adaptive per-video frame budgets — cut VLM inference cost by 75% and frame count by 48% on the Pitt Ads Dataset while improving topic accuracy to 86.0% (+1.6 pp). Paper under review at ACM Multimedia 2026.
-
Summer Research Assistant (Robotics)
May — Aug 2025Wabash College · Math & CS Dept, advised by Dr. Qixin Deng
Autonomous navigation fusing GPS, computer vision, and LiDAR with multi-sensor fusion and emergency protocols. 97% obstacle detection accuracy and ~2 m positional precision in outdoor testing.
-
Machine Learning Research Intern
Aug 2024 — Apr 2025Wabash College
Built a U-Net + ResNet model for facial-emotion recognition across 135 compound classes; collaborated with cybersecurity and HCI partners on behavioural-threat detection. Presented at IEEE ICAIC 2025.
05 : Awards
- Presidential International Scholarship Wabash College
- J. Crawford Polley Mathematical Writing Prize May 2025
- CalHacks-12 Winner Berkeley · 2025
- UAP CyberSiege-25 Winner 2025
06 : Coursework
computer science
- Data Structures & Algorithms
- Machine Learning
- Data Science
- Machine Organization
- Compiler Design
- Autonomous Robotics
- Functional Programming
- Physics-Informed ML
mathematics
- Multivariable Calculus
- Linear Algebra
- Differential Equations
- Numerical Methods
- Stochastic Simulation