Ayobami

Full-Stack (Mobile & Web) and AI Engineer

I build production-grade AI/ML systems optimized for mobile devices and resource-constrained environments.

Current Work

Building EchoLLM — a desktop application that provides seamless access to state-of-the-art models from frontier labs (Claude, GPT, Gemini) with a privacy-first approach. Built with Flutter and available on the Snap Store.

Actively developing:

• Offline model support via llama.cpp integration
• Custom model distillation pipelines for mobile deployment
• Quantization techniques to run advanced models on phones and laptops

Tech Stack

---

Featured Projects

Rubik-YOLO — Real-Time Rubik's Cube Solver Vision System

A specialized computer vision model that accurately detects and identifies Rubik's Cube face colors and orientations in real-time, designed as the visual foundation for automated cube-solving systems.

What It Does:

• Detects and classifies all 6 cube face colors (Blue, Green, Orange, Red, White, Yellow) simultaneously
• Handles non-standard cube orientations using Oriented Bounding Boxes (OBB)
• Operates robustly under varying lighting conditions and camera angles
• Processes frames at 30+ FPS on mid-range Android devices

Technical Implementation:

• Fine-tuned YOLOv8-nano on a self-curated, diverse Rubik's Cube dataset
• Custom annotations created in Label Studio with support for rotated bounding boxes
• Achieved ~0.95 mAP over 72 training epochs with early stopping
• Converted PyTorch weights to ONNX format for mobile NPU acceleration
• Built inference pipeline supporting live camera, static images, and video processing

Use Case: Serves as the perception layer for Rubik's Cube solving applications, enabling real-time state detection needed for algorithmic solving.

EchoLLM — Privacy-First Mobile LLM Runtime

On-device inference engine enabling offline LLM execution on mobile platforms, with optional cloud model access via BYOK (Bring Your Own Key) usage model.

Technical Highlights:

• Distilled reasoning capabilities from DeepSeek-R1 (671B) into a 1.7B Qwen model, preserving 95% of benchmark performance
• Engineered 4-bit quantization pipeline reducing model size to <1.2GB while maintaining sub-200ms first-token latency
• Built high-performance C++ FFI bridge between llama.cpp and Flutter, achieving 20 tokens/s on consumer hardware

Impact: Enables privacy-preserving AI assistants that operate completely offline—no cloud APIs, no telemetry, no compromises.