- Published on
An experiment in fine-tuning small, local language models (Qwen 1.5B) to reason in a telegraphic, compressed style, saving up to 74% of reasoning tokens and speed up inference without relying on verbose prompts.
Local-ai
All Posts
- rag (4)
- how-to (3)
- visualization (3)
- agents (3)
- webgpu (3)
- wasm (3)
- transformersjs (3)
local-ai (3)
- ml (2)
- short_read (2)
- llm (2)
- onnx (2)
- python (2)
- retrieval (2)
- maths (1)
- news (1)
- sovereign-ai (1)
- data-center (1)
- compute (1)
- india (1)
- gpu (1)
- infrastructure (1)
- ai-policy (1)
- browser-ai (1)
- pglite (1)
- webllm (1)
- pgvector (1)
- mlx (1)
- sft (1)
- reasoning (1)
- qwen (1)
- deepseek (1)
- voice (1)
- markdown (1)
- pdf (1)
- typesetting (1)
- serverless (1)
- aws (1)
- azure (1)
- gcp (1)
- sandboxing (1)
- Published on
A deep dive into exporting and optimizing AI4Bharat's IndicTrans2 models for local, offline translation in browser runtimes via ONNX, WebAssembly, and WebGPU.- Published on
A technical deep-dive into building WebVoice Studio, a fully local, zero-server voice AI workbench running entirely in the browser using WebGPU, WASM, Silero VAD, and custom model runner adapters.