01
Adaptive layer-wise quantization
4–8 bit integer precision per layer, chosen dynamically from gradient-norm history — replacing FedSparQ's fixed FP16 scheme with something that actually responds to how much each layer is changing.
M.Tech · IIT Jodhpur . Researcher
Compressing intelligence
to fit on the edge.
I'm Tabrej Alam — a federated learning researcher building communication-efficient training systems for resource-constrained devices. My current work, EF‑AdapSparseQ, pushes adaptive quantization and error-feedback past fixed-precision baselines under heavy non‑IID conditions.
8.92
CGPA · M.Tech SIOT
7
FL methods benchmarked
0.021
EF buffer norm @ round 15
live render — client ↔ server gradient exchange
scroll
// 01 — about
Background & trajectory
I'm currently pursuing my M.Tech in Electrical Engineering (Sensors & IoT) at the Indian Institute of Technology, Jodhpur, ranked 2nd in the ongoing batch. My thesis sits at the intersection of federated learning and model compression — specifically, how to quantize and sparsify gradient updates so that edge devices with limited bandwidth and compute can still train collaboratively without leaking raw data or collapsing under non‑IID conditions.
Before this, I completed my B.Tech in Computer Science & Engineering at HNB Garhwal University, where I built full‑stack web applications and explored applied machine learning — a foundation that now shows up in how I engineer FL simulation pipelines rather than just theorize about them.
I'm preparing my current results — EF‑AdapSparseQ — for presentation to my thesis supervisor and for submission to IEEE Access, framed around accuracy‑per‑bit gains under severe non‑IID data splits (α = 0.1).
2025 —
M.Tech (SIOT), IIT Jodhpur · CGPA 8.92 · Rank 2
2025 —
Qualified GATE
2021–25 —
B.Tech CSE, HNB Garhwal University · CGPA 7.67
2020 —
Class XII, Jeevan Public School, Motihari
2018 —
Class X, St. Michael's Academy, Bettiah
// 02 — research
Federated learning, compressed
My thesis proposes EF‑AdapSparseQ, extending FedSparQ with three concrete innovations aimed at communication-efficient, robust FL on edge hardware.
02
Gradient-norm bit scheduling
A coarse‑to‑fine bit-width schedule driven by gradient-norm trends, so early noisy rounds spend fewer bits and later, more meaningful updates get more precision.
03
Per-client error-feedback buffers
Each client retains its own EF buffer to carry forward quantization error, preventing gradient drift under heavy non‑IID partitioning.
Validated result
Error-feedback buffer norm converges near zero (~0.021 at round 15) versus FedSparQ's persistent oscillation (~0.23) — empirically supporting the convergence theorem behind the method.
- LeNet‑5 on CIFAR‑10
- Dirichlet non‑IID, α = 0.1 / 0.5 / 1.0
- vs. FedAvg, QSGD, Top‑K, FedPAQ, FedProx, FedSparQ
// 03 — build log
Things I've shipped
Motor speed control via computer vision
Real-time gesture and color-based motor control. Hand detection at 120ms, color detection at 85ms — 95% gesture accuracy, 92% color-detection accuracy.
PythonOpenCVReal-time CV
E-mail spam detection system
Classical ML pipeline using Python and Scikit-learn, reaching 92% classification accuracy on spam vs. legitimate mail.
PythonScikit-learnNLP
Event management system
Web application for managing events, bookings and schedules end‑to‑end, improving event scheduling efficiency by roughly 40%.
HTML/CSSJavaScriptMySQL
GUI scientific calculator
A desktop scientific calculator built with a Python GUI toolkit — clean operator precedence handling and a no-frills interface.
PythonTkinter
// 04 — stack
Tooling that compiles, trains, and ships
Languages
CC++Python
JavaVerilogSystemVerilog
Web
HTMLCSSJavaScript
React.jsNode.jsExpress.js
Data
MySQLMongoDB
Tools & platforms
GitGitHubVS Code
VivadoEDA PlaygroundKeil
MATLABGoogle ColabPyTorch
research interests —
Federated Learning for IoT
/
Timing Analysis of RTL
// 05 — signal
Achievements
2nd
Currently ranked 2nd in the ongoing M.Tech SIOT programme at IIT Jodhpur
✓
Qualified GATE 2025
⟁
Graph Theory Programming Camp Certificate — AlgoUniversity
// 06 — connect
Let's exchange gradients.
Let's exchange gradients.
(or just talk research.)
Open to discussions on federated learning, edge AI, model compression, or full-stack collaboration. Reach out directly — no intermediary server required.
email
m25eet009@iitj.ac.in
phone
+91 6204 765522
linkedin
/in/tabrej-alam0178
current location
IIT Jodhpur, Rajasthan, India