Progressive Knowledge Distillation Framework for Military Vehicle Classification on Edge Devices
A ResNet-18 student beats its ResNet-50 teacher at 89.38% accuracy, and the 1.6 MB TFLite model runs on Android in 100-300 ms.
Akshar Nath Gorain
AI Researcher & Systems Engineer. I compress heavy neural networks into efficient on-device models and ship them across firmware, phones and the web.
- Lines: VisionVision 8 stopsDetection, segmentation and recognition in images and video.
- Lines: SignalsSignals 5 stopsRadio, audio, ECG and camera-pulse signals.
- Lines: EdgeEdge 8 stopsModels and firmware that run on the device itself.
- Lines: ProductProduct 8 stopsApps, web services and backends that people use.
Research
An accepted IEEE paper, a Top 5 poster and two studies under way. Full record on ORCID.
EdgeRF: Sub-Megabyte Drone Controller Identification at Low SNR
99.89% across 17 RC drone controllers with one SNR-agnostic model (prior best 95.62%), in a 0.59 MB INT8 model at about 2 ms per CPU inference. It also re-evaluates the DroneRF dataset for train/test leakage.
Liver and Tumour Segmentation from CT Scans
Top 5 Best Poster. SLIC-superpixel post-processing on a fine-tuned Swin-UNet lifts tumour precision from 0.875 to 0.982 on LiTS (131 CT volumes), with liver Dice 0.86. Benchmarked against SAM2, MedSAM, MobileSAM and PicoSAM.
Multiscript Handwritten Character Segmentation & Recognition
Lazy Dataset training, YOLOv8-MSER triple-layer detection, and ConvNeXt-EfficientNet ensemble with script-aware filtering. International Rank #1 (Score: 0.755).
Projects
Viewing All 7 Projects
Seven projects. Pick a line to light up the ones that run on it, and open a row for the details.
Milestones
Click any milestone for 3D technical deep-dive ✦| Year | Milestone | From | Lines | View |
|---|---|---|---|---|
| 2026 | COMSYS International Hackathon 7, during the 7th International Conference of COMSYS Champion 🏆 | Škoda Auto University / Jadavpur Univ. | Lines: Vision | Inspect → |
| 2026 | Paper accepted, IEEE SPACE Accepted Paper 🛰️ | IEEE Aerospace & Electronic Systems Society | Lines: Vision, Edge | Inspect → |
| 2026 | Top 5 Best Poster, IndiSetu Kolkata Top 5 Poster 🏅 | IndiSetu Healthcare AI Summit | Lines: Vision | Inspect → |
| 2026 | 3rd Prize, HexaFalls 2 (Accio Relevance) 3rd Prize 🥉 | JIS University | Lines: Product | Inspect → |
| 2026 | 3rd Prize, Made in JIS (Celeb Edition) Startup Award 🚀 | JIS Group Incubation Centre | Lines: Product | Inspect → |
| 2026 | Co-authored "Edge-AI Chip" Article Magazine Article 📰 | Tech Darshi Magazine | Lines: Edge | Inspect → |
| 2026 | CURAA joins the incubation cohort AIC Incubated 🐾 | Atal Incubation Centre | Lines: Signals, Edge, Product | Inspect → |
| 2026 | AI-ML Virtual Internship (Grade O) Grade O 🎓 | AICTE and EduSkills | Inspect → | |
| 2025-26 | Cloud & GenAI certifications Certified ☁️ | AWS, Google Cloud, Meta, UPenn | Inspect → |
About
I'm a B.Tech CSE student (3rd year) at Techno India University, Kolkata (2024-2028). I work at the intersection of Computer Vision, Knowledge Distillation, and Edge AI — compressing heavy neural networks (ResNet18 to MobileNetV2) into efficient on-device models. International Champion at COMSYS Hackathon 7, co-author at IEEE SPACE 2026, Top 5 Best Poster at IndiSetu 2026.
Route so far
Bokaro
Guru Gobind Singh Public School, 2021
CBSE, Class X. 84.60%
Bokaro
Guru Gobind Singh Public School, 2023
CBSE, Class XII. 77.83%
Kolkata
Techno India University, 2024-2028
B.Tech, Computer Science and Engineering. CGPA 7.69 (Sem 4)
Core Engineering & Research Pillars
Lines: Vision Vision & Deep Learning
Neural compression, dense medical segmentation & multiscript recognition.
Core Work: Progressive Knowledge Distillation (IEEE SPACE), SLIC Swin-UNet (IndiSetu Top 5), YOLOv8-MSER (COMSYS Champion #1).
Lines: Signals Signal Intelligence & Biometrics
Physiological vitals extraction, anti-spoofing audio & radio frequency telemetry.
Core Work: On-device camera rPPG 60 FPS BVP extraction (HeartSense), sub-megabyte DroneRF ID (EdgeRF), distilled AASIST voice armor.
Lines: Edge Edge AI & Hardware Systems
Low-compute edge quantization, bare-metal microcontrollers & DSP firmware.
Core Work: INT8 quantization under 2MB RAM, ESP32-S3 FreeRTOS embedded runtimes, TinyML on-chip inference without cloud dependency.
Lines: Product Systems & Quant Architectures
Ultra-low latency execution engines, real-world platforms & cryptographic verification.
Core Work: Rust sub-millisecond crypto derivatives terminal (<90MB RAM), UN SDG 2 food rescue engine (ShareBite), decentralized ledger audit.
Languages: Python, C/C++, Rust, Kotlin, Dart, Solidity, SQL, Bash, PHP