# AshnaAI Research

> Technical notes and development research from AshnaAI on agents, routing, artifacts, connectors, and governed execution.

Hub: https://www.ashna.ai/research

## Research articles

- [Ashna-X1 and task-aware model routing](https://www.ashna.ai/research/ashna-x1-task-aware-model-routing): Ashna-X1 is the default catalog model on AshnaAI, but it does not replace every provider endpoint. Task-aware routing means matching tools, context limits, and reasoning controls to the job instead of sending every request to one chat model.
- [From research to deck: AshnaAI’s agentic presentation pipeline](https://www.ashna.ai/research/research-to-deck-agentic-presentation-pipeline): AshnaAI does not generate a PowerPoint from one prompt. A create orchestrator runs request analysis, research, planning, image generation, and code. Updates use a separate orchestrator so an existing deck is not rebuilt from scratch.
- [TallyPrime desktop bridge: GST reads and controlled writes](https://www.ashna.ai/research/tallyprime-desktop-bridge-for-gst-and-controlled-writes): AshnaAI’s Tally tools are schema-only on the server. Execution happens on the user’s desktop. GST paths refuse to infer tax treatment, and write paths prefer review over silent ledger creation. The desktop app is required.
- [No-code agent specs and tool permissions](https://www.ashna.ai/research/no-code-agent-spec-and-tool-permissions): AshnaAI’s agent builder writes a structured spec, not a free-text prompt dump. Tools are attached by internal IDs. Permissions that need OAuth or extra user data are requested in schema instead of being invented by the model.
- [Client-side browser control versus server RPA](https://www.ashna.ai/research/client-side-browser-control-vs-server-rpa): AshnaAI controls the user’s own browser through an extension or desktop broker. It is not a hosted RPA farm. Actions are a closed list, every run needs an idempotency key, and secrets go to requestUserInput instead of the page.
- [Voice and telephony agents for live calls](https://www.ashna.ai/research/voice-and-telephony-agents-for-live-calls): AshnaAI can place and receive live calls through LiveKit telephony plus Twilio or Telnyx numbers. Inbound routing attaches a SIP trunk or voice webhook. Numbers are filtered for voice capability before purchase.
- [Grounding agents in user files](https://www.ashna.ai/research/grounding-agents-in-user-files): AshnaAI grounds answers in files the user already uploaded. Embeddings are queued in the background, linked to the chat before ingestion starts, and can be restarted if they go stale. The product retrieves from those files; it does not train a new model on them.
- [Human-in-the-loop before irreversible actions](https://www.ashna.ai/research/human-in-the-loop-before-irreversible-actions): AshnaAI stops an agent before irreversible work by using requestUserInput and review gates. The input tool has four kinds: question, choice, file, and acknowledgement. Tally writes use the same idea as Need-review instead of silent creates.