private AI · CPA firms
Client files in, first drafts out
An agent that reads client documents, sorts receipts and invoices, drafts client emails and answers questions from your own guidance. Tax files stay on a box in your office.
- Data stays in your building
- You own the model and the agent
- No subscription required
What the agent does
- 01
Read and sort client documents
Invoices, receipts and statements sorted, with the key fields pulled into a file your software can import.
- 02
Draft client emails and memos
Missing-document requests, deadline reminders and explanations, in your tone, for your staff to send.
- 03
Answer from your procedures
Staff questions answered from your firm's procedures and past work papers, with the source.
How we prepare it
The box and the open models are the same ones anyone can buy: a Mac Studio or a GB10-class machine, Qwen, Llama, Gemma or Mistral. What differs is how they are prepared for your work, and that preparation is what we do.
- 01
Audit your data
We look at your documents, recordings and the tasks that eat your week, and agree on one task for the pilot.
- 02
Build a test set
From your own documents we write the questions the model has to answer and the answers it has to give. Everything after is measured against that set.
- 03
Tune the model, train the agent
We fine-tune an open model on your documents and teach the agent your workflows: your templates, your archive, your forms.
- 04
Check before handover
The model has to pass your test set before we hand it over. You see the score, not a demo.
Then you own it: the box, the model, the agent, the test set and the admin keys. Support and updates are there if you want them.
We have no ready-made model for any field yet. We are taking a few pilots for 2027, and each model is prepared from that client's own data.
What an answer looks like
The agent answers from your files and names the document and section it used, so a person can check it in seconds.
You: Which March receipts are missing for the quarterly filing?
Two: the office rent receipt and the fuel card statement. Both show on the bank statement with no matching document.
↳ Source: Bank_Statement_March.pdf · lines 14 and 27
Who builds it
A small team that has built AI video products and AI agents for years. Everything below is public and linked.
AI video since 2014
The first Life2Film shipped in 2014 as a web service. The rebuilt app scores every frame and finds scene changes on the device, on a Rust engine of 11 crates, and has been in the App Store since August 2026.
Life2Film →Agents, in the open
solo-factory (57 skills and 2 agents for Claude Code), a coding agent written in Rust and an MCP server that searches code and past sessions. All open source.
Open source →Hackathons
2nd place at the Supabase Hackathon, 3rd of 78 at a San Francisco cyberdefense hackathon with a real-time video tool, and a judge at Cal Hacks 13.0 in October 2026.
Write-ups →Founder: Rustam Salavatov · About the company
What the law says
-
A preparer who knowingly or recklessly discloses a client's return information, or uses it for anything other than preparing the return, commits a misdemeanor unless IRS regulations allow it.
The regulations allow sharing with outside providers only for return preparation and only inside the United States; other disclosures need the client's signed consent. A model on your own hardware keeps return information inside the firm. 26 CFR 301.7216-2.
-
Tax preparation firms count as financial institutions under the rule. They must pick service providers that can safeguard client data, require it by contract and check them periodically.
Every AI vendor that receives client files becomes one of those providers. A local model adds none.
Civil penalty under 26 U.S.C. §6713: $250 for each prohibited disclosure or use, up to $10,000 a calendar year. eCFR, 26 CFR 301.7216-1(a).
For information, not legal advice. Checked against the official texts on 5 November 2026.
Three ways to start
Model
An open model tuned on your documents that answers with sources. Runs on hardware you already have, if it is strong enough.
Model + agent
Plus an agent that does the work in your tools: drafts from your templates, searches your archive, fills your forms.
Turnkey box
Everything on a compact box we install in your office and hand over to you.
Questions
Does it file returns?
No. It prepares and drafts; your staff review, sign and file.
Which software does it work with?
Anything that imports and exports files: we connect through the formats your software already supports. Tell us your stack in the form.
Does any of our data go to a cloud?
No. The model runs on the box in your office, and the training happens there too. The box can work with no internet access at all if your policy asks for that.
Why on-premise?
Some data should not leave the building: privileged client files, patient records, tax files. Keeping the model in your office is the simplest way to keep that data under your control. We do not sell a compliance certification; your own counsel decides what your obligations require.
Which models?
Open models whose licences allow business use, such as Qwen, Llama, Gemma or Mistral. We pick one that fits your box and your work, and can swap it when a better one ships.
Who owns it?
You do: the hardware, the tuned model, the agent and its settings. Nothing stops working if you stop paying us. Support and updates are optional.
How much does it cost?
It depends on the size and the work. Tell us what you need and we will send a quote. We are taking a small number of pilot clients and agree the pilot scope with each.
What do we need to provide?
One person who knows the work, the documents or video the AI should use, and a place for the box with power and a network port.
What a pilot looks like
- 01
Call
We talk through your documents and pick one task.
- 02
Data check
We look at the files that task needs and tell you what is usable.
- 03
Test set
Questions and expected answers, written from your own documents.
- 04
First version on the box
The tuned model and the agent running in your office.
- 05
Review
We run the test set together and decide what comes next.
How long each step takes depends on your data, so we agree the scope before we start.
Ask for a pilot
Tell us what the AI should work with. A person reads it and replies by e-mail within two working days.
Sold, installed and supported in the US by SuperDuperAI, Corp., a Delaware public benefit corporation.
Mac mini and Mac Studio are trademarks of Apple Inc. SuperDuperAI is not affiliated with Apple.