Avishek Shrabon

AI Researcher — Federated & Representation Learning

+880 1618 569 306 avishek@avishekkerketa.dev avishekkerketa.dev github.com/Avishek7777 linkedin.com/in/avishek-shrabon

Research

HSBN-FL: Hierarchical Stochastic Bottleneck Networks for Heterogeneous Federated Learning

Independent Research — In Progress, Core Experiments Complete

2025 — Present

Targets heterogeneous federated learning: letting institutions with differing data and model architectures — hospitals, NGOs, research labs — train collaboratively without sharing raw data or model weights
Replaces weight exchange with latent-space communication — clients share compressed representations rather than parameters, natively supporting architectural heterogeneity across participants
Structures the server as a two-level hierarchy: a Transformer adapter (Z1) attending over all client representations at once, and an MLP apex (Z2) compressing the pooled output into a global abstract state; top-down feedback returns to each client as a soft alignment signal
Evaluated on CIFAR-100 across 20 heterogeneous clients under Non-IID Dirichlet partitioning; held accuracy variance under 0.7% across heterogeneity levels

Hierarchical Stochastic Bottleneck Networks: Conditions for Representational Abstraction in Deep Hierarchies

Manuscript Under External Review — Targeting IEEE Conference

2025

Asks under what conditions deep networks form a genuine hierarchy of abstractions rather than encoding redundant information across layers — a question at the intersection of information theory and deep learning
Identifies three necessary mechanisms — per-level objectives, bandwidth-limited stochastic channels (a KL penalty), and bidirectional message passing — under which each level learns a strictly more compressed representation than the one below
Verified empirically on CIFAR-10 and CIFAR-100 using Centered Kernel Alignment and intrinsic dimensionality; ablations isolate two qualitatively distinct failure modes

Adaptive Federated Learning with Heterogeneous Data and Client Distributions

Undergraduate Thesis — AIUB Academic Repository · Grade: A+

2024 — 2025

Designed a pipeline pairing gradient-based client clustering with cross-architecture knowledge distillation (ResNet and ViT) to handle non-IID data in federated settings; learned the domain independently in roughly six weeks
Supervised by Debajyoti Karmaker (formerly AIUB; now Lecturer at RMIT & AIH, Australia)

Education

Bachelor of Science in Computer Science & Engineering

American International University Bangladesh (AIUB) — Dhaka, Bangladesh

Jan 2022 — Dec 2025 CGPA 3.74 / 4.00

Held a merit scholarship continuously from the third semester through graduation
Awarded the Dean's Award (Spring 2022–23) for academic excellence — CGPA above 3.8
Relevant coursework: Computational Statistics & Probability (A+), Algorithms (A), Software Requirements Engineering (A+)

Experience

Private Tutor — Sciences & Mathematics

Self-employed — Dhaka, Bangladesh

Mid 2022 — Present

Tutored 3–4 students concurrently across Grades 8–12 in Physics, Chemistry, Biology, Mathematics, Higher Mathematics, and ICT — sustained alongside a full-time degree and independent research

Projects

1000 Missionary Movement Bangladesh — Training Platform

Full-Stack Web Application — Live Deployment, Nationwide NGO

2025 — Present

Worked with NGO leadership and programme coordinators to replace a manual, paper-based intake process with a digital platform, coordinating directly with non-technical stakeholders throughout
Built with Next.js, Prisma, and Auth.js; handles applicant intake, role-based access, and programme administration for hundreds of participants nationwide

Church Event Ledger & Ticket Reconciliation System

Full-Stack Software — Live Deployment, 5 Consecutive Events

2025

Spotted an operational bottleneck in event logistics and shipped a working tool in three days with the event team
Cut end-of-night accounting from 15–20 minutes to under 5 — a 75% reduction — now standard practice across subsequent events

Proximity Sensor Aid for the Visually Impaired

Hardware — Embedded Systems & Assistive Technology

2023

Built a wearable obstacle-detection aid using an Arduino Nano, HC-SR04 ultrasonic sensor, and an experimental Time-of-Flight sensor; evaluated its limitations and proposed a path toward better range and reliability

Darkness-Triggered Street Light Automation

Hardware — Electronics & Infrastructure

2023

Designed an LDR-based system that switches relay-controlled street lights on at dusk, motivated by energy efficiency in public infrastructure

Skills

Languages Python, JavaScript, TypeScript, SQL; C++ and C# (familiar)

AI / ML PyTorch, scikit-learn, Hugging Face, NumPy, Pandas, Seaborn, Matplotlib

Web & Frameworks React, Next.js, Nest.js, Node.js, Vite, HTML, CSS/SCSS

Databases PostgreSQL, MongoDB, SQLite

Tools Git, GitHub, VS Code, Google Colab, Kaggle, Linux

Research Federated Learning, Knowledge Distillation, Distributed Systems, Information Theory, LLMs, NLP

Awards & Activities

Dean's Award

American International University Bangladesh

Spring 2022–23

Runners-Up — Annual Volleyball Tournament

American International University Bangladesh

2024

Volunteer — Church Community Services

Event coordination, software deployment, and community engagement

Ongoing

References

Anjir Ahmed Chowdhury

PhD Researcher, Intelligent Data and Systems Lab — University of Houston, USA

aachowd4@cougarnet.uh.edu

S M Abdullah Shafi

Lecturer, Special Assistant (OSA) — American International University Bangladesh

shafi@aiub.edu

Debajyoti Karmaker

Lecturer (RMIT & AIH), Australia — Undergraduate Thesis Supervisor

Further contact details available upon request.