parkky

everything glued in so far

Projects

Notes from the lab — voice agents, language models & multimodal experiments.

SLM — Marathi Language Model

GitHub ↗

Dec 2024

Pretrained an 84M-parameter GPT-2-inspired Marathi LM from scratch (6 layers, 6 heads, 384 dim). Built a custom 32K Marathi tokenizer and a 66K+ story dataset by translating TinyStories.

PyTorchGPT-2TokenizerPretraining

OpenBee — Offline Voice Assistant

GitHub ↗

Mar 2026

Fully offline voice AI (Speech → LLM → Voice) with zero cloud dependency. Whisper + Gemma 1B + Kokoro on LiveKit. TTFT < 80ms, TTS < 500ms on consumer hardware, with a React control dashboard.

WhisperGemmaKokoroLiveKitReact

MemorySearch — Semantic Image Search

GitHub ↗

Jan 2026

Modular image retrieval using BLIP captioning + configurable embeddings (Qwen0.6B, GTE, Gemma). Config-driven, incremental indexing and multi-model similarity ranking without full re-indexing.

BLIPEmbeddingsRetrievalMultimodal

LocalMind — Local Agentic RAG

GitHub ↗

Jul 2025

Open-source, fully local conversational AI with LlamaIndex + LangGraph. Agentic RAG via dual tools (docs + web) and persistent memory. Q5-M quantized Jan-nano LLM (~60% VRAM savings).

LlamaIndexLangGraphRAGFastAPI

Alice — Home Surveillance System

GitHub ↗

Feb — May 2025

Low-latency AI security using QwenVL-2.5 4B for real-time threat detection, quantized for ~43% VRAM reduction. Evaluation loop cut false positives ~35%; LiveKit + Twilio SIP calls homeowners with context-aware alerts.

QwenVLQuantizationLiveKitTwilio

Draupadi — AI Safety App

Nov 2024

Real-time distress detection app (React Native + TensorFlow + Twilio) that detects voice cues like "help" or screams. GPS tracking + auto-SMS alerts and an "I'm Safe" mode for user control.

React NativeTensorFlowTwilioGPS