I am an AI / ML Engineer specializing in On-Device Machine Learning, Android Systems (C++/NDK), and High-Performance Edge Runtimes. My engineering focus revolves around executing large language models (LLMs) and computer vision runtimes 100% locally on edge devices with zero cloud telemetry, low RAM footprints, and real-time inference latency.
- AI / ML Engineering: Deep integration with LiteRT (TFLite), MediaPipe GenAI, and native
llama.cppruntimes for quantization, NPU/GPU hardware acceleration, and dynamic context windows. - Software Systems Engineering: Production-grade Kotlin, C++20, JNI bridges, CMake cross-compilation, and memory-safe Android NDK development.
- Product & UI Engineering: Modern Jetpack Compose design systems, real-time Haze glassmorphism blur effects, custom graphics rendering, and gesture navigation.
- Interactive Audio Computing: Low-latency real-time DSP, pitch & amplitude dynamic tracking, and hands-free voice interface engines.
Open To: Senior Software Engineer Roles | On-Device AI / Mobile ML Engineering | Native Android Systems Collaboration
-7C3AED?style=for-the-badge&logo=tensorflow&logoColor=white){kind=icon}
| Domain | Proficiency | Core Technologies & Details |
|---|---|---|
| On-Device LLM Inference | Expert | Local execution of 1B–4B parameter models (DeepSeek R1, Gemma 3, Qwen 2.5, Llama 3) on mobile SOCs. |
| Native Runtime Acceleration | Advanced | LiteRT-LM CPU/GPU backends, MediaPipe GenAI task pipeline, and native llama.cpp C++ JNI integration. |
| Quantization & Memory Optimization | Advanced | INT4 / INT8 quantization alignment, KV-cache memory limits, thermal profiling, and NPU delegate mapping. |
| Private & Offline Architecture | Expert | Zero-telemetry sandbox storage, local prompt session caching, gated Hugging Face token local encryption. |
🤖 Solus — On-Device Private AI Assistant (Click to Expand)
Solus is an open-source, enterprise-grade, 100% offline Android application that executes state-of-the-art Generative AI models directly on user devices with zero server latency or cloud tracking.
| Metric | Details |
|---|---|
| Stack | Kotlin, C++20, Android NDK, CMake, Jetpack Compose, LiteRT, MediaPipe GenAI, llama.cpp |
| Scale | Multi-model catalog supporting 1B–4B LLM & Vision parameters |
| Performance | Streamed real-time generation (~24 t/s on Snapdragon 8 Gen 3) with TTFT < 150ms |
| Security | 100% local sandbox storage, zero analytics telemetry, encrypted HF read token memory |
| Impact | Fully functional production release (v1.5.0) available via direct APK release |
| Repository | github.com/ShounakPatra/Solus |
Solus features a dynamic glassmorphism UI built with Jetpack Compose and real blur algorithms, single-row telemetry metrics (t/s speed, TTFT, temperature, RAM), and automatic NPU/GPU hardware acceleration.
🎮 Vocex-Run — Voice-Controlled Android Game (Click to Expand)
Vocex-Run is an interactive, real-time voice-controlled endless runner game for Android where dynamic microphone input replaces traditional touch controls.
| Metric | Details |
|---|---|
| Stack | Kotlin, Android SDK, Real-Time Audio DSP, Canvas Rendering, Android AudioRecord API |
| Scale | Multi-mode environment scaling (Silent, Normal, Rage Mode) |
| Performance | Low-latency audio sampling loop (< 20ms buffer latency) |
| Security | Local audio processing buffer discarded immediately after pitch extraction |
| Impact | Production release v5.1 with persistent score storage and optimized shadow pipelines |
| Repository | github.com/ShounakPatra/vocex-run |
Vocex-Run extracts real-time amplitude and pitch features from live microphone buffers to control character movement, jump triggers, and velocity scaling without physical touch inputs.
Self-Employed / Open Source Contributor • 2025 — Present
Architecting, developing, and deploying high-performance native Android applications and AI runtimes.
- Designed and shipped Solus, integrating LiteRT-LM, MediaPipe GenAI, and
llama.cppnative C++ bindings for 100% offline LLM inference. - Implemented real-time hardware telemetry tracking (tokens/sec, TTFT, device thermal state, free RAM) to prevent memory crashes on low-end hardware.
- Authored custom Jetpack Compose UI systems featuring dynamic theme switching, Haze blur shaders, and full gesture navigation support.
- Built Vocex-Run, an audio-driven Android game featuring low-latency DSP algorithms for real-time voice intensity control.
Kotlin C++ Android NDK LiteRT MediaPipe llama.cpp Jetpack Compose CMake
| Recognition | Details |
|---|---|
| 🚀 Solus v1.5.0 Release | Shipped production-ready offline AI Android app with dynamic model themes & telemetry guard. |
| 🎮 Vocex-Run v5.1 Release | Shipped stable, low-latency audio-driven Android game with custom score persistence. |
| 🟢 100% Passing Android CI | Maintained automated GitHub Actions workflows for testing, linting, and build verification. |
Certified Cloud Foundations & Architecture Fundamentals
Java SE Programming & Object-Oriented Software Design
Elite Certification in Programming in Java
Data Structures and Algorithms in C++
CCNA Introduction to Networks & Security
Learning:
- Advanced Quantization (AWQ, GPTQ, GGUF K-Quants)
- Vulkan GPU Shader Acceleration for Mobile AI
- Agentic Tool-Use Execution Runtimes
Building:
- Solus v5.0.0 (llama.cpp C++ JNI & Native GGUF Engine)
- High-Performance Android Audio DSP & Camera Vision Extensions
Exploring:
- Low-Latency Local RAG Embeddings (Vector Databases on Android)
- Android System Server Architecture & Memory Guard
Open To:
- Senior On-Device AI Engineering Roles
- High-Impact Open Source Software Collaboration
