everything glued in so far
Notes from the lab — voice agents, language models & multimodal experiments.
my first side hustle · live
a little atelier for love letters
You write slowly, choose the paper and the wax, and send one quiet link. They break the seal; the words unfold.
This Ganpati season, three people paid for it. No ads. Someone told someone, which is the oldest kind of letter there is.
Three sales, zero ads, Ganpati season 2026
p.s. built for one reader first. she read it.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.
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.
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.
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).
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.
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.
the next quest — someday soon
The bloom gets the photographs. The agents do the work.
Next: an AI revolution in the field. The bulb is in the ground.
