# Data — Language, speech and reasoning APIs

One OpenAI-compatible API for chat, vision, OCR, transcription, speech synthesis, translation and retrieval — running on dedicated hardware, with models hosted in the service itself.

- Year: 2026
- Category: AI
- Stack: litellm, caddy, python, fastapi, uvicorn, websocket, llamacpp, pytorch, cuda, huggingface, nvidia, librosa, scipy, ffmpeg, tesseract, easyocr, pymupdf, searxng, postgresql, redis, sqlite, docker, systemd
- Live: https://console.data.larsima.com

### One API for every language task

Chat, vision, OCR, transcription, speech synthesis, translation and retrieval — behind one standard interface. The models run locally, on dedicated hardware.

### What you get

The models run locally on dedicated hardware, so sensitive text and audio do not have to be sent to an outside service to be processed. For a team working with legal files and recorded conversations, that is the whole point.

Under the hood

### What it is built from

Data is built in layers. At the edge, Caddy handles HTTPS and rate limiting, and behind it a LiteLLM gateway serves the OpenAI-compatible interface, with API keys, per-model access and spend tracking. The service itself is written in Python with FastAPI and Uvicorn, and live transcription is streamed over WebSocket. The models run on a dedicated NVIDIA Blackwell GPU with 96 GB of memory: quantised language models through llama.cpp, the rest on PyTorch 2 with CUDA 12.8 and Hugging Face Transformers. Speech synthesis with a neural audio codec, fine-tuned for Persian with LoRA, produces the voices, and librosa, SciPy and FFmpeg handle the audio work. Documents go through two independent OCR engines, Tesseract and EasyOCR, whose results are settled by consensus voting, with PyMuPDF reading PDFs. Retrieval combines BM25 with embeddings, and web search runs through SearXNG. Data lives in PostgreSQL, cached responses in Redis and local indexes in SQLite, with Docker and systemd keeping the service running.


---

Source: https://larsima.com/zh/projects/data-ai-apis
Organisation: LARSIMA (شرکت فن‌آوران توسعه لار سیما), registered in Iran, no. 318515, since 2007-12-02.