AI Playground

Role: Director & Lead Engineer | Tech: Kotlin Multiplatform, Compose Multiplatform, C++20 / NDK (JNI), llama.cpp v0.7.0, Google AI Edge LiteRT, Apple MLX, Vulkan/OpenCL/Metal, SQLDelight

An offline-first, on-device multimodal AI workstation for Android and iOS that turns mobile devices into self-contained inference engines. Powered by Kotlin Multiplatform, Compose Multiplatform, native C++20 NDK (llama.cpp v0.7.0), and Apple MLX Swift / Metal GPU, it executes open-source LLMs (GGUF, Safetensors, MediaPipe), synthesizes on-device diffusion images (Bonsai Image Ternary 4B / FLUX.2-klein), parses 25+ document formats locally, and delivers desktop-class generation speed with zero cloud subscriptions and 100% data privacy.

Key Achievements:

  • Dual-Model Speculative Turbo Engine: Implemented speculative drafting with companion draft models (e.g., Qwen 2.5 0.5B + Llama 3.2 1B), dynamic draft context capping (std::min(2048u, ...)), RAM admission guardrails (isSpeculativeTurboEligible), and KV cache session pre-warming to double decoding throughput while preventing VRAM exhaustion.
  • Multi-Runtime Ecosystem: Engineered a unified runtime orchestrating llama.cpp v0.7.0 (with modular dynamic GPU feature delivery for Vulkan/OpenCL and ARM KleidiAI CPU backend), Google AI Edge LiteRT-LM (.litertlm / .task), LiteRT 2.3.0 Multi-Graph Image Generation (Bonsai 4B), and Apple MLX / Metal GPU on iOS.
  • On-Device Diffusion & Bonsai 4B Image Studio: Built an on-device image synthesis studio supporting 16 curated models with 4-step photorealistic generation, anti-hallucination presets, and W3C PNG tEXt chunk metadata attribution.
  • Zero-Dependency KMP Document Engine: Developed a pure Kotlin Multiplatform document processing engine extracting semantic structured text across 25+ file formats (PDF, DOCX, XLSX, PPTX, CSV, JSON, Markdown, RTF, OpenDocument) using pure RFC 1951 Deflate/ZIP decompression with zero external cloud APIs.
  • Crash-Resilient Hardware Isolation: Designed a defensive fork() child process sandbox with a non-blocking 3000ms watchdog polling loop and kill(pid, SIGKILL) fallback for native Vulkan driver probing, preventing vendor driver GPU SIGSEGV crashes or driver deadlocks from taking down the host app.
  • 100% Offline Activity Ledger: Built a privacy-first local analytics ledger (SQLDelight schema v10) tracking usage milestones, tokens, and active streaks with strictly zero external network calls (INV-OFFLINE-01).
  • Streaming Fault-Tolerance: Architected periodic mid-generation SQLite state persistence (STREAM_PERSIST_INTERVAL_MS = 2500ms) with automatic recovery of interrupted assistant turns (PARTIAL_FAILED with preserved output text) upon process restart.
  • Cross-Platform Parity & Multimodal Voice: Delivered full Android & iOS parity featuring offline continuous voice dictation (Apple Speech framework + Android STT), sentence-level multilingual TTS, real-time context token headroom gauges, and an adaptive two-pane layout.

Live Links: AI Playground on Google Play | GitHub Repository | Deep-Dive Engineering Series

VelociHue

Role: Director & Lead Engineer | Tech: Kotlin Multiplatform, Compose Multiplatform, Kotlin 2.4.0, AGP 9.2.1, Target SDK 37 (Android 17), Cloud Firestore (Schema v5), Android TV

A fast-paced, high-framerate arcade action game engineered in Kotlin Multiplatform and Compose Multiplatform, shipping natively across Android, iOS (SwiftUI host + shared KMP framework), and Android TV (Leanback & D-pad integration). Built on modern toolchains (Kotlin 2.4.0, AGP 9.2.1, Target SDK 37), it features an endless color-gate survival loop, dynamic physics, particle effects, multi-tiered progression, and conflict-free cross-device cloud synchronization.

Key Highlights:

  • Shared Pure-Kotlin Game Engine: Architected a shared deterministic game loop (GameEngine, CollisionDetector, StatsManager, AchievementManager) running at high refresh rates (60–120 FPS), managing color-matching physics, projectiles, power-ups (Shield, Slow-Mo, Magnet), and revive flows across Android and iOS.
  • Conflict-Free Cloud Sync Engine (Schema v5): Designed a cross-device sync engine backed by Cloud Firestore (CloudGameData) with 3-way union merge strategies that safely synchronize player progression, currency, cosmetics, daily stats, and weekly rolling recaps without data regressions.
  • 29-Achievement Multi-Tiered Engine: Implemented a comprehensive achievement catalog covering 1–5,000 player levels, 100k diamonds, gameplay feats, cosmetic collection milestones, 7–30 day retention streaks, and skill tiers (Untouchable, Pacifist, Sharpshooter, Pure Run) evaluated by a single pure-function derived unlock engine (AchievementUnlockComputation) with retroactive upgrade backfill.
  • Sensory & Accessibility Suite: Integrated a photo-sensitive and reduced-motion safety mode (dampening screen shake, flashes, and slow-mo vignettes), granular SFX/music volume sliders with snap-to-detents and haptic clicks, and dynamic haptic intensity scaling.
  • Native Android TV & 10-Foot Experience: Built dedicated television deployment with Leanback manifest declarations, D-pad navigation focus states, and TV-optimized layouts.
  • Instant Compose Share Cards: Engineered a run-summary share card composable captured directly into high-resolution PNGs via Compose GraphicsLayer.toImageBitmap() and shared seamlessly via Android FileProvider.
  • Multi-Flavor Build Modernization: Decoupled distribution architecture supporting internal, staging, and production dimensions with automated buildSrc version governance and iOS xcconfig synchronization.

Live Link: VelociHue on Play Store

Digital Commerce Platform

Role: Lead Android Engineer | Tech: Kotlin, Jetpack Compose, MVVM, Clean Architecture

Building sophisticated digital commerce solutions for global markets with seamless payment integration, inventory management, and real-time synchronization across multiple platforms.

Key Achievements:

  • Architected scalable mobile solutions handling millions of transactions
  • Implemented adaptive layouts for diverse device configurations
  • Reduced app load time by 40% through performance optimization

Hardware Integration Module

Role: Senior Developer | Tech: Kotlin, Android NDK, Hardware APIs, Low-Latency IPC

Developed comprehensive hardware integration layer enabling seamless connectivity with POS systems and peripherals.

Key Achievements:

  • Integrated 27+ hardware devices (printers, scanners, payment terminals)
  • Achieved 99.8% connection reliability
  • Reduced hardware communication latency by 60%
  • Created reusable abstraction layer for future integrations

Open Source Contributions

GitHub: @DearDhruv

Active contributor to Android and Kotlin Multiplatform open source ecosystems with a focus on architecture patterns, on-device AI runtimes, testing utilities, and developer tools.

Looking for more details? Feel free to explore the Blog for technical deep-dives on these projects and related topics.