# Capabilities

The atomic capabilities behind our locally-deployed AI systems: large language models (selection, evaluation, local deployment, fine-tuning), speech (TTS/STT), retrieval (RAG, rerank, GraphRAG), and custom AI agent development.

---

# Atomic Capabilities

Every system we ship is composed from four layers, and we build and tune each one
ourselves. That is what separates us from a cloud resale: we can reach into the
**large language model**, the **speech pipeline**, and the **retrieval layer** and
adjust them for your task, your data, and your hardware.

## Large language models

We start with model selection and evaluation — benchmarking candidates against the
actual work, not a leaderboard. The chosen model is then **deployed locally** on
your infrastructure and, where it helps, **fine-tuned** on your domain data and
quantized to run efficiently on your GPUs.

## Speech

The voice layer covers both directions: **speech-to-text** (STT) and
**text-to-speech** (TTS), supporting English, Mandarin, and Cantonese. We tune the
pipeline for low, predictable latency and a natural voice, including domain
vocabulary and accent work.

## Retrieval

Retrieval-augmented generation (**RAG**) grounds the model in your own documents. We
add **reranking** for relevance and **graph-enhanced search (GraphRAG)** so the
model can reason across connected information — and every answer is traceable back
to a source.

## Agent development

Finally, we compose the model, voice, and retrieval layers into a **custom AI
agent** that fits your workflow: answering, booking, routing, or assisting your
team, end to end, and deployed locally.