# Sourodyuti Biswas Sanyal

> AI researcher and builder. I train vision and signal models, then ship them on firmware, phones and the web. Co-founder of CURAA and Curawave LLP.

I'm a B.Tech CSE student at Techno India University, Kolkata (2024-2028). Most of my work starts as a model and ends as something people can hold: a collar, a phone app, a web service. I write the firmware, the app, the backend and the model myself, and I co-founded CURAA and Curawave LLP along the way.

Location: Durgapur and Kolkata, India. Open to research roles, internships and collaborations. Email: sourodyuti.biswas.sanyal.14@gmail.com

## Links

- [github](https://github.com/Sourodyuti)
- [linkedin](https://www.linkedin.com/in/sourodyutibs/)
- [kaggle](https://www.kaggle.com/sourodyutibiswas)
- [orcid](https://orcid.org/0009-0003-3809-9448)
- [instagram](https://www.instagram.com/sourodyutibiswas/)
- [email](mailto:sourodyuti.biswas.sanyal.14@gmail.com)
- [curaa](https://curaapets.com)
- [resume](https://www.sourodyuti.me/resume.pdf)

## Research

### Progressive Knowledge Distillation Framework for Military Vehicle Classification on Edge Devices
Accepted. IEEE SPACE, Bengaluru, 2026. Lines: Vision, Edge

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.

### EdgeRF: Sub-Megabyte Drone Controller Identification at Low SNR
In preparation. IEEE RadarConf, 2027. Lines: Signals, Edge

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
Poster. IndiSetu Kolkata, 2026. Lines: Vision

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.

### AI Stereoscopic 2D to 3D Video Conversion
In progress. Independent research, 2025-now. Lines: Vision

A depth-estimation pipeline (MiDaS, Depth Anything V2) that turns ordinary video into side-by-side 3D for passive displays, optimised with TensorRT for a 6 GB laptop GPU.

## Projects

### CURAA (2024-now)
Technical co-founder. AIC incubated. Lines: Signals, Edge, Product. [Join the waitlist](https://curaapets.com)

An AI collar that listens to what pets can't say.

My startup. A smart dog collar that maps a dog's sounds and movement to behaviour: coughing, scratching, barking, changes in sleep and activity. The collar's microphone and motion sensor stream to the paired phone, which does the processing. It is a health collar, not a GPS tracker. I own the whole stack: ESP8266 and ESP32-S3 firmware, a TinyML bio-acoustic engine under 200 KB, the Flutter app and the Cloudflare backend. Selected for the Atal Incubation Centre. Co-founded with Aritra Biswas and Krishnendu Dhar.

Stack: ESP32, TinyML, Flutter, React, Cloudflare Workers

### HeartSense (2026)
Co-builder, team of four. Shipped. Lines: Vision, Signals, Edge, Product. [GitHub](https://github.com/workforakng/heartsense)

Your phone camera can read your pulse.

Contactless vital signs: look into a phone camera for 30 seconds and get heart rate, SpO₂, respiration, HRV and stress, with no wearable and no internet permission. CameraX exposure and white-balance lock, 478-point MediaPipe face regions, and a POS/CHROM, Butterworth and adaptive Kalman DSP core running on Android through Chaquopy. 1.64 BPM mean absolute error on UBFC-rPPG, and on-device PDF reports in under 100 ms.

Stack: rPPG, MediaPipe, CameraX, Kotlin, Chaquopy

### VaaniKavach (2026)
Sole architect. In development. Lines: Signals, Edge, Product

A bodyguard for your phone calls.

India lost ₹22,495 crore to cyber fraud in 2025, and AI voice cloning is the fastest-growing weapon. VaaniKavach is an on-device sentinel that listens to live calls, detects synthetic voices with a distilled anti-spoofing model, flags scam pressure (urgency, fear, authority) in Bengali and Hindi, and warns before money moves. Every flag becomes a hash-chained evidence pack designed to stand up in court.

Stack: AASIST to TFLite, Kotlin, Whisper, SHA-256 evidence chain

### Cardiosense (2026)
Curawave LLP, for Cardiosense AI. Client work. Lines: Signals, Edge, Product

Firmware and an Android app for a portable cardiac-screening prototype.

Client work through my company, Curawave LLP: embedded firmware, on-device ECG signal analysis and an Android app for Cardiosense AI's portable cardiac-screening research prototype, delivered under formal change control.

Stack: Embedded firmware, ECG analysis, Android

### Distillation Lab (2024-2026)
Researcher. Published. Lines: Vision, Edge

Shrinking big vision models until they fit in a pocket.

The research programme behind my IEEE SPACE paper: progressive teacher, assistant and student distillation that squeezes ResNet and ViT-scale models into small deployable students with accuracy nearly intact. Tested on CIFAR-10, MedMNIST and a 6-class military vehicle dataset.

Stack: PyTorch, Attention transfer, ViT, ONNX

### Melt-Pool Vision (2026)
Researcher. IEEE report. Lines: Vision, Edge

Teaching cameras to watch metal being 3D-printed.

Real-time melt-pool segmentation and classification for Directed Energy Deposition additive manufacturing. I trained Fast-SCNN on high-speed coaxial imagery, benchmarked architectures for latency and FPS, and exported to ONNX for edge deployment. Written up as an IEEE-format research report.

Stack: PyTorch, Fast-SCNN, ONNX, LaTeX

### COMSYS Hackathon 7 (2026)
ML lead, Team Vic-Trio. Champion. Lines: Vision

First place in multi-script handwriting recognition.

At Jadavpur University: segmenting and recognising handwritten characters across several scripts in degraded scanned documents. Our entry won, ranked #1 on the public leaderboard at 0.766. The pipeline was YOLO detection, a ConvNeXt ensemble classifier and test-time augmentation, trained on twin Kaggle T4s.

Stack: YOLO v8 / v11, ConvNeXt ensemble, Test-time augmentation

### SehatKhata (2026)
Co-builder, hackathon team. 36-hour build. Lines: Vision, Product

India's medical records live in plastic bags. We replace the bag.

An offline-first, voice-first PWA that photographs a family's paper prescriptions and turns them into a health timeline. It checks for duplicate therapy, interactions and allergies with deterministic rules (no LLM in the decision path), warns while a prescription is being typed, shares consent-scoped QR briefs with doctors, explains in Hindi by voice, and has an emergency card and caregiver mode. Built in 36 hours as a pnpm and Turborepo monorepo: 7 apps, 10 packages, 3 workers.

Stack: Fastify, Prisma, NVIDIA NIM vision, NATS JetStream, PWA

### Legilimens (2025-2026)
Creator. Open source. Lines: Product. [GitHub](https://github.com/Sourodyuti/Legilimens)

A confusion radar for live classrooms.

Students press one button the moment they get lost. The system captures the lecture context, retrieves the right explanation from a hybrid BM25 and vector knowledge vault, and replies with a personalised voice explanation within seconds. Professors get a live view of exactly where the class lost the thread. FastAPI backend, Next.js frontend, fully on-prem. 2nd Runner-Up, Best Actian Usage at HexaFalls 2.0.

Stack: FastAPI, WebSockets, sqlite-vec, Next.js

### ResumeIQ AI (2026)
Full-stack builder. Live. Lines: Product. [Open the app](https://resumeiq-ai-9qc.pages.dev)

Reads your résumé the way a recruiter would.

A live SaaS: a hybrid ATS rule engine plus an LLM critique pipeline that parses any résumé, scores career readiness on five weighted factors, maps missing skills and predicts role fit. A Next.js 15 monorepo with a Hono API on Cloudflare Workers.

Stack: Next.js 15, Hono, Cloudflare Workers, LLM pipeline

### Delta Terminal (2026)
Sole architect. Active. Lines: Product

A terminal for crypto derivatives, written in Rust.

A Rust-native terminal for BTC and ETH futures and options: real-time neural-ensemble inference, full options chains with Greeks, IV and OI surfaces, a fee-aware signal engine, and a headless CLI built so AI agents can operate it. About 90 MB of RAM, sub-millisecond internals.

Stack: Rust, Neural ensemble, Options Greeks, WebSockets

## Milestones

- 2026: Champion, COMSYS Hackathon 7, Jadavpur University
- 2026: Paper accepted, IEEE SPACE, IEEE
- 2026: Top 5 Best Poster, IndiSetu Kolkata, IndiSetu
- 2026: 2nd Runner-Up, Best Actian Usage, HexaFalls 2.0, JIS University
- 2026: CURAA joins the incubation cohort, Atal Incubation Centre
- 2024: Grade O, AI/ML internship, AICTE and EduSkills
- 2024: Leadership bootcamp, BIT Mesra
- 2023-25: Cloud and GenAI certifications, Oracle, Google Cloud, Unstop

## Education

- 2022: ICSE, Class X, Dr. Graham's Homes, Kalimpong (83.2% (English + best 4))
- 2024: ISC, Class XII, St. Michael's School, Durgapur (77.5% (English + best 3))
- 2024-2028: B.Tech, Computer Science and Engineering, Techno India University, Kolkata (SGPA 7.98 / 10 (3rd semester, latest))

## Tools

- Vision: PyTorch, TensorFlow, OpenCV, YOLO, Swin / ViT, SAM2, MediaPipe
- Signals: DSP filtering, rPPG (POS / CHROM), RF classification, AASIST anti-spoofing, Whisper
- Edge: ESP32 / ESP8266, TinyML, TFLite Micro, FreeRTOS, Knowledge distillation, INT8, ONNX, TensorRT, BLE
- Product: TypeScript, React / Next.js, Flutter, Kotlin, Hono, FastAPI, Cloudflare Workers, PostgreSQL
- Languages: Python, TypeScript, Rust, C++, Kotlin, Go, Dart