Prathmesh Nikam
AI & Machine Learning Developer — B.Tech AI & Data Science (Final Year)
I build and ship production-grade AI systems end-to-end: generative AI pipelines, LLM fine-tuning, speech synthesis, and computer vision — optimized for constrained consumer hardware (6GB VRAM GPUs, Raspberry Pi edge devices). Every project here is independently built, documented, and deployed.
🚀 Featured Projects
LLM & RAG
Gemma 3-1B Fine-Tuning
Fine-tuned Google’s Gemma 3-1B for Python code generation using QLoRA on a 6GB GPU. Achieved ~60% memory reduction vs full fine-tuning. Published on Hugging Face as GGUF.
RAGv2 — Offline Retrieval-Augmented Generation
Privacy-first local RAG with parent-child chunking, FAISS vector search, and IBM Granite 350M LLM — all running under 2s latency on consumer hardware with zero cloud dependency.
Speech & Audio AI
Qwen3-TTS GPU Suite
GPU-accelerated markdown-to-speech using Alibaba’s Qwen3-TTS. Voice cloning + 9 preset speakers, SDPA attention, bfloat16 — all running locally on an RTX 3050.
Whisper Dictate
Push-to-talk speech-to-text daemon for Wayland using OpenAI’s Whisper. Hold a hotkey, speak, release — transcription appears in any focused application. 100% local.
md2voice
Documentation-to-audio pipeline that uses Cerebras AI to rewrite technical docs into conversational text, then synthesizes speech via edge-tts or Piper TTS.
Computer Vision & Multimodal
Driver Drowsiness Detection
Real-time drowsiness detection using custom YOLOv11m-cls classification with dual-threshold temporal logic. Achieves 96.4% mAP@50 with multi-modal visual/audio/OS alerts.
Local Semantic Media Search
Privacy-first semantic search for local image and video collections. Uses Qwen3-VL-Embedding-2B (2048-dim) with ChromaDB for sub-100ms vector lookups. No cloud APIs.
Guides & Tooling
llama.cpp Guide
Comprehensive guide for running LLMs locally on Windows using llama.cpp with various hardware backends (CUDA, Vulkan, Metal, CPU).
🛠️ Skills
| Domain | Technologies |
|---|---|
| LLM & Training | PyTorch, Unsloth, LoRA/QLoRA, PEFT, bitsandbytes, TRL, Hugging Face Transformers, GGUF |
| Speech & Audio | Qwen3-TTS, edge-tts, Piper TTS, ONNX Runtime, Cerebras API |
| Computer Vision | YOLOv11 (Ultralytics), OpenCV, InsightFace (ArcFace), MediaPipe |
| Vector Search & RAG | ChromaDB, FAISS, llama-cpp-python, Qwen3-Embedding |
| Web & Deployment | Next.js 15, React 19, TypeScript, Firebase, FastAPI, Docker, Vercel, Tailwind CSS |
| Edge & Embedded | Raspberry Pi 4/5, model quantization/pruning, ARM inference |