Full-Stack & Systems Engineer | Mobile Core, Distributed APIs & Applied Edge ML
Akshar Nath Gorain

Akshar Nath Gorain

Building low-latency mobile platforms, real-time monitoring tools, and automated system utilities. Focused on applied ML and production-grade architectures.

Undergraduate Researcher at Techno India University & Co-builder at CURAA. International Champion at COMSYS 2026. Translating high-dimensional machine learning and low-level Linux/mobile kernels into reliable client platforms.

View Resume (PDF) GitHub Profile LinkedIn Email / Contact
Kolkata, West Bengal, India · Available for high-impact roles & research
// about me

The Mission.

B.Tech CSE, 3rd year, Techno India University — bridging cutting-edge AI research with production architectures.

🧠

AI Researcher

Knowledge Distillation and Computer Vision. INT8 Quantization, YOLOv8 pipelines, MSER ensembles. 🏆 COMSYS 2026 International Champion — Škoda Auto University, Czech Republic.

🐾

Mobile & ML @ CURAA

Cross-platform Flutter app and intelligent service pipelines for CURAWAVE's pet-care ecosystem. Native 60fps rendering. Co-built with Sourodyuti Biswas Sanyal.

⚡

Full-Stack Engineer

ShareBite (Next.js 14 + PostgreSQL), VeriFi blockchain auth, Delta Terminal algo-trading system. Systems that ship and scale.

PythonRustPyTorch KotlinJetpack ComposeFlutter Next.js 14FastAPIC/C++ SolidityPostgreSQLAWS DockerRAG & MCPBash/Termux YOLOv8Edge AI / INT8Gemini API
// experience

The Timeline.

From first commit to research labs, startups, and national hackathon podiums.

Undergraduate AI Researcher
Aug 2024 – Present
Techno India University · Computer Vision
  • Innovating core pipelines for low-compute deployment — Knowledge Distillation & INT8 Quantization.
  • Pioneering diagnostic medical imaging compression for edge-device runtime optimizations.
PyTorchPythonEdge AI
Mobile App & ML Engineer
Nov 2025 – Present
CURAA · CURAWAVE LLP
  • Architected cross-platform Flutter ecosystem orchestrating veterinary, grooming, and ML service pipelines.
  • Engineered custom hardware rendering at native 60fps. Collaborated with Sourodyuti on CURAA & Delta Terminal.
FlutterDartMobile MLNode.js
Enterprise AI Engineering Intern
Oct 2026 – Present
AICTE – EduSkills · Virtual (Cohort 19)
  • Building enterprise Retrieval-Augmented Generation (RAG) pipelines, high-dimensional vector search indexing, and Model Context Protocol (MCP) multi-agent workflows.
RAGVector SearchMCP AgentsAICTE
AI-ML Virtual Internship
Aug 2026 – Oct 2026
AICTE – EduSkills (Google for Developers) · Virtual
  • 8-week applied AI-ML program — model building, evaluation, and deployment fundamentals. Completed with Grade O (Outstanding).
PythonMachine LearningAICTE
AI Automation Intern
June 2026 – July 2026
AICTE – IBM SkillsBuild · Virtual
  • Orchestrating intelligent automation workflows, AI agents, and enterprise AI integrations.
AI AgentsAutomationIBM
Cybersecurity Intern
June 2026 – Aug 2026
AICTE – EduSkills (Palo Alto Networks) · Virtual
  • Enterprise network security, threat intelligence platforms, and SOAR automation.
CybersecurityPalo AltoSOAR
Government ML Intern
Past
AICTE · Remote
  • Developed ML models predicting 6 major water pollutants using historical station data.
PythonMLAICTE
// selected works

Systems I've Shipped.

From hackathon-winning CV pipelines to production full-stack platforms.

Open Source

HeartSense — Contactless Vital Monitoring

CMOS Camera ➔ CHROM/POS DSP ➔ LiteRT ➔ 5 Vitals @ 60 FPS
Architecture & Constraints: Extracted real-time Photoplethysmography (PPG) waveforms directly from mobile CMOS sensors using CameraX and lightweight Google LiteRT inference pipelines.
Engineering Impact: Optimized frame-buffer processing in C++/Kotlin to maintain 60 FPS inference without thermal throttling on mid-range Android hardware (zero cloud roundtrips, 100% on-device).
LiteRTKotlin / ComposeOpenCVSignal DSP (FFT)CameraX
GitHub Repo Architecture Spec
Production App

CURAA — Mobile Core & Service Orchestration

Flutter 60fps UI ➔ Offline SQLite Sync ➔ Payment Webhooks
Architecture & Constraints: Engineered the core client orchestration layer integrating scheduled asynchronous vet bookings, payment gateway webhooks, and state caching for unstable 4G networks.
Engineering Impact: Reduced cold-start load times by 40% via lazy asset resolution and modular state stores; delivered native 60 FPS transitions across 5 microservices.
Flutter / DartRESTful APIsState ManagementPostgreSQLMobile ML
Live Platform Co-built w/ Sourodyuti
Open Source

Linux & Termux System Automation

Boot Watcher ➔ Heartbeat Daemon ➔ Self-Healing SSH Tunnel
Architecture & Constraints: Headless bash/daemon automation engine running under Termux/Linux environments for autonomous boot recovery, remote tunnel healing, and continuous telemetry logging (vnstat, CPU/RAM buffers).
Engineering Impact: Achieved 99.9% uptime on remote headless Android nodes; instant automated Telegram alerting on dynamic WAN IP changes and unauthorized login spikes.
BashLinux SignalsTermux APISSH TunnelingNetwork Sockets
GitHub Repo System Breakdown
IEEE Prep

EdgeRF — RF Drone Identification

Raw I/Q Spectrograms ➔ INT8 Quantized CNN ➔ 0.59 MB Payload
Architecture & Constraints: Formulated lightweight INT8 quantized neural classification over high-frequency RF bursts across 17 commercial drone remote controllers.
Engineering Impact: Achieved 99.89% classification accuracy within a 0.59 MB footprint and ~2ms CPU inference latency; rigorously proved and fixed train/test leakage in published DroneRF benchmarks.
PyTorchEdge AIINT8 QuantizationSignal DSPEmbedded C++
Research Artifacts
Open Source

VaaniKavach — Voice Clone Defense

Real-time Audio Stream ➔ AASIST Distilled TFLite ➔ SHA-256 Ledger
Architecture & Constraints: Designed zero-latency telephony defense isolating deepfake synthetic voice synthesis in incoming calls using distilled AASIST graph architectures.
Engineering Impact: Runs 100% locally on Android without off-device audio transmission; provides multilingual coercion keyword recognition and tamper-evident SHA-256 evidence logging.
TFLiteAudio MLSecurityAndroid NDKForensics
GitHub Profile
Private Core

Delta Terminal — Algo Execution Core

Live WebSocket Orderbooks ➔ Five-Gate Filter ➔ <1ms Tick Routing
Architecture & Constraints: Low-latency derivatives trading terminal in Rust & Python executing live BTC/ETH options strategies under strict memory bounds (<90MB RAM).
Engineering Impact: Sub-millisecond tick dispatch with zero GC pauses; implemented multi-alpha interference models and headless remote CLI monitoring. Co-engineered with Sourodyuti.
RustPythonFlask / SocketsAsyncIODerivatives
Proprietary / Private
Live Production

ShareBite Platform — Food Rescue Engine

Next.js 14 App Router ➔ PostgreSQL / Prisma ➔ Real-Time Ledger
Architecture & Constraints: Real-time surplus inventory distribution engine aligned with UN SDG 2 (Zero Hunger), featuring strict role-based access control and expiry tracking.
Engineering Impact: Zero-latency donation matching via optimistic UI mutations; resilient atomic state transactions preventing double-claim race conditions.
Next.js 14PostgreSQLPrisma ORMNextAuthTailwind
Live Demo
Top 5 Poster Award

Liver Tumor Segmentation — Swin-UNet

CT Scans ➔ Swin-UNet Embeddings ➔ SLIC Superpixel Boundary Refinement
Architecture & Constraints: Swin Transformer encoder-decoder paired with Simple Linear Iterative Clustering (SLIC) superpixel boundary consensus on the LiTS medical benchmark.
Engineering Impact: Outperformed MedSAM and SAM2-Tiny baseline models on fine edge lesion detection; honored with Top 5 Best Poster at IndiSetu 2026 AI Healthcare Summit.
Swin-UNetSLICMedSAMPyTorchMedical Imaging
Poster Abstract
Open Source

Vision Knowledge Distillation

ResNet-18 (Teacher) ➔ KL Divergence Loss ➔ MobileNetV2 (Student)
Architecture & Constraints: Logit-level distillation pipeline transferring dark knowledge from heavier CNNs into lightweight edge backbones under automated mixed precision (AMP).
Engineering Impact: Achieved 4.8x compression ratio with less than 1.4% accuracy degradation on CIFAR-10; integrated WandB telemetry and ONNX export.
PyTorchKnowledge DistillationEdge AIONNX
GitHub Repo
Hardware + App

Cardiosense — Edge ECG Platform

AD8232 Sensor ➔ ESP32-S3 (Rust) ➔ BLE Stream ➔ Compose Client
Architecture & Constraints: Battery-powered 3-lead ECG screening platform with 100% on-device analysis and zero cloud reliance.
Engineering Impact: Embedded Rust firmware reading analog buffers from dual AD8232 chips; real-time low-energy Bluetooth packet streaming to a modern Kotlin/Jetpack Compose interface.
ESP32-S3Embedded RustBLEKotlinBiomedical
Physical Hardware Prototype
Open Source

VeriFi — Blockchain Auth

Document Hash (SHA-256) ➔ Solidity Contract ➔ On-Chain Proof
Architecture & Constraints: Ethereum dApp for verifiable academic credential notarization and attendance records using SHA-256 digital signature proofs.
Engineering Impact: Deployed gas-optimized smart contracts ensuring non-repudiation and zero-tamper record verification across institutions.
SolidityEthereumWeb3.jsNext.js
GitHub Repo
// engineering architecture

Technical Deep Dives.

How I reason through bottlenecks, system latency, and edge compute trade-offs.

SYSTEM DESIGN · EDGE DSP

Sub-Millisecond PPG Signal Processing: Eliminating Motion Artifacts via POS/CHROM & Adaptive Kalman Filtering

Extracting blood volumetric changes from standard RGB CMOS phone cameras introduces severe motion noise, uneven sensor exposure, and non-linear lighting shifts. The core challenge: maintaining sub-bpm accuracy while processing 60 FPS video streams on resource-constrained mobile CPUs without causing thermal throttling or dropping camera frames.

[CameraX 60 FPS RAW YUV] ↓ Zero-copy frame lock [Facial ROI Spatial Downsampling] → 120x120px skin cluster ↓ POS (Plane-Orthogonal-to-Skin) & CHROM algorithms [BVP Waveform Extraction] → Dynamic range normalized ↓ 2nd-order Butterworth (0.75–3.0 Hz) + Kalman Filter [LiteRT Peak Inversion Engine] → HR, SpO&sub2;, HRV, Stress @ 60 FPS
  • Zero-Copy Memory Buffer: Directly queried Android hardware ImageProxy planes via native C++ buffers, avoiding RGB raster conversion overhead and saving ~14ms per frame.
  • Motion Robustness: Projected RGB color signals onto a plane orthogonal to the skin tone vector (POS), isolating genuine pulse-induced capillary volume changes from ambient flicker.
  • Compute Efficiency: Replaced heavy 3D ConvNets with lightweight temporal DSP filters + a quantized 1.2MB LiteRT model, preserving 60 FPS on sub-$150 Android devices.
SYSTEMS ENGINEERING · HEADLESS LINUX

Autonomous Daemon Healing: Building Zero-Root Telemetry & Remote Tunnels under Mobile Linux

Running persistent automation on mobile Linux (Termux) without root privileges faces aggressive Android battery killers (Doze mode, Phantom Process Killer, and OOM-killer). This architecture constructs a crash-resilient watchdog engine that maintains reverse SSH tunnels, dynamically binds network sockets, and streams hardware telemetry.

[Android Boot Receiver (Termux:Boot)] ↓ Spawn daemon with flock mutex [Supervisor Daemon (POSIX Shell)] ↔ SIGTERM / SIGHUP Trapped ↓ Periodic heartbeat (10s ring buffer) [Tunnel State Inspector] → Socket probe (netstat / /proc/net/tcp) ↓ Failure detected? Tear down & exponential backoff [Autossh Reverse Tunnel + Telegram Dispatcher] → 99.9% Remote Node Uptime
  • PID Lock & Orphan Reaping: Enforced single-instance execution via Linux kernel flock descriptors and automated cleanup of zombie child processes.
  • Self-Healing Tunnel Logic: Monitored TCP keepalive roundtrips instead of relying solely on process state, triggering reconnects immediately upon carrier IP rotation.
  • Proactive Telemetry: Compressed rolling system logs (RAM buffers, vnstat network bandwidth, and device temperature) and dispatched asynchronous Telegram webhooks during security anomalies.
// all work

Everything I've Built.

Open source and private — AI pipelines, system automation, blockchain, and more.

❤️
HeartSensepublic

Contactless rPPG vital-signs screening — 5 vitals, 100% on-device at 60 FPS.

📡
EdgeRFpublic

RF drone-controller identification — 99.89% across 17 RC controllers, 0.59 MB INT8.

🛡️
VaaniKavachpublic

AI deepfake voice clone scam defense — on-device AASIST TFLite + audio forensics.

🫀
Cardiosenseprivate

Battery-powered 3-lead ECG platform — ESP32-S3 Rust firmware + BLE Compose client.

🎓
Legilimensprivate

Real-time classroom confusion radar — hybrid BM25 + vector voice explanations.

🍽️
ShareBitepublic

Full-stack food rescue platform — UN SDG 2, real-time ledger, cross-role logging.

🐾
CURAAprivate

Flutter app for CURAWAVE pet-care ecosystem — 5-microservice architecture.

🔬
Liver Tumor Segpublic

SLIC superpixel + Swin-UNet medical CT segmentation (IndiSetu 2026).

⚡
Delta Terminalprivate

Institutional algo-trading — Five-Gate theta engine, multi-alpha interference. Co-built with Sourodyuti.

🧠
Vision KDpublic

PyTorch teacher-student distillation — ResNet18 → MobileNetV2, CIFAR-10.

🔗
VeriFipublic

On-chain document verification — SHA-256 + Ethereum smart contracts.

💧
Water Quality AIpublic

ML models predicting 6 water pollutants — AICTE Gov. internship project.

🏥
Jīva Healthprivate

Healthcare record digitization — AI-powered OCR/NLP extraction for India.

🤖
Physics Solver AIpublic

Step-by-step physics problem solver powered by Gemini via Express backend.

🌦️
FlintWeatherAIpublic

Lightweight AI weather assistant aggregating free weather provider APIs.

💻
SSH Termux Autopublic

IP monitoring, failed login alerts, battery notifications via Telegram.

🎰
Prediction Backendprivate

Flask API with worker processes, simulator blueprint, and pattern recognition.

🩸
Blood Donation BDMSpublic

Web app for donor registration, inventory tracking, and request management.

📞
Phonebook Managerpublic

Contact manager with C backend and PHP frontend — CRUD with file storage.

19 projects · 11 public · 8 private
// recognition

Highlights & Publications.

🏆
International Champion — COMSYS Hackathon 7
7th Intl. Conference on Frontiers in Computing & Systems · Škoda Auto University, Czech Republic · Team Vic-Trio (Akshar, Sourodyuti & Subhodip) · Rank #1 · Score 0.755
🛰️
IEEE SPACE 2026 Presenter & Co-Author
3rd Space, Aerospace and Defence Conference, Bangalore · Progressive Knowledge Distillation for Military Vehicles on Edge Devices
🥉
HexaFalls 2 — 3rd Prize ($200 USD)
Accio Relevance Hackathon · MLH Official Season · JIS University · Team Traction
🥉
Made in JIS, Celeb Edition 2026 — 3rd Prize
Team CURAA · Kolkata · Aug 2026
🏅
Top 5 Best Poster — IndiSetu 2026
AI in Healthcare Summit, Kolkata · SLIC Superpixel Swin-UNet Liver Tumor Segmentation
☁️
AWS Academy Graduate
Data Engineering Certified · Dec 2025
📰
Tech Darshi Coverage (Sept 2026)
Featured Edge-AI Research Coverage · Jadavpur University, Kolkata
// contact

Let's Work Together.

Open to research collaborations, internships, and building things that matter.

Contact me directly at workforakng@gmail.com or drop your info here.

I'll never share your data with anyone else. Pinky promise! 🤙