I specialize in Knowledge Distillation, On-Device Inference, and Computer Vision. My focus is building systems that run heavy compute workloads efficiently on the Edge (iOS/Android/Linux).
| Domain | Technologies |
|---|---|
| AI & ML | PyTorch, TensorFlow, ONNX Runtime, llama.cpp, LoRA/QLoRA, Hugging Face |
| Mobile & Systems | C++, Dart (Flutter), Swift, Bash, Linux Kernel, Docker |
| Data & Backend | Python, FastAPI, PostgreSQL, Vector Databases (Chroma/Pinecone) |
| Infrastructure | AWS (SageMaker), GCP, GitHub Actions (CI/CD) |
EchoLLM | On-Device Inference Engine
A privacy-first mobile LLM runtime capable of running 1B+ parameter models offline.
- Distillation: Built a pipeline to distill reasoning capabilities from DeepSeek-R1 into a 1.7B parameter Qwen model.
- Optimization: Manually quantized models to 4-bit GGUF, reducing memory footprint to <1.2GB while maintaining 95% of reasoning benchmarks.
- Systems: Implemented a C++ FFI bridge to link the underlying inference engine directly to the Flutter UI, achieving 20 tokens/s on consumer mobile hardware.
An NLP engine that maps abstract human sentiment to Spotify Audio Features.
- Architecture: Replaced standard keyword search with a semantic vector mapping system (User Prompt -> Valence/Energy/Danceability Vectors).
- Integration: deeply integrated with the Spotify Web API to generate algorithmic playlists based on "Vibe" rather than Genre.
- Stack: Flutter, Python (NLP Middleware), Spotify OAuth 2.0.
👁️ Rubik-YOLO | Real-Time Mobile Computer Vision
High-performance object detection system for 3D spatial objects.
- CV Pipeline: Trained a custom YOLOv8 model using Oriented Bounding Boxes (OBB) to track Rubik's cubes in dynamic lighting.
- Deployment: Converted PyTorch weights to ONNX for optimized cross-platform execution on mobile NPU/CPUs.
- Data Ops: Built an active learning loop using Label Studio to correct model drift.
- Bible_Extractor (ETL): A T5-based NLP pipeline for normalizing and structuring unstructured archaic text data.
- Jemma: CLI-based developer tooling for automating data validation workflows.
- Symbolic Distillation: Experimenting with distilling "Chain of Thought" reasoning from 70B models into <3B mobile-class models.
- Edge Quantization: Benchmarking perplexity degradation in GGUF vs. AWQ formats for mobile NPUs.
- GitHub: ThatLinuxGuyYouKnow
- LinkedIn: LinkedIn