Data stays in the company
The model runs on your computer or server. The contents of mails, and client and patient data, do not reach any AI provider.
Agent LMD · a local decision model from Tomcat Flow
A model that returns decisions instead of text, computed on your own hardware
Agent LMD reads a case and picks one of the options your own software offers it, with a probability for each. Software can act on an answer like that directly: fill the form, set the priority, route the case. The model runs on your computer, so nothing reaches an AI provider. We deploy it together with software written for your process, because that is what turns the model's decisions into work in your systems.
* Measured on purpose-built test sites with fictional data, 29 September 2026. How we measured
How it works
The same steps for every case. Each decision comes with how confident the agent is.
- 01
Reads the case
A mail, a message or a report. The agent works out what it is about and which system to use.
- 02
Decides
Every decision is a choice among a few options with a visible confidence, 97% for instance. You set the threshold: below it the agent does not act alone, it hands the case to a person.
- 03
Acts
Fills the forms and clicks in your systems, step by step, in the same windows your team works in. No API integration needed where a system has none.
- 04
Built for your process
There is one model, but the questions we ask it are written for your process. That is not cosmetic: it is the part that turns a general-purpose model into something that finishes your cases.
- 05
You set the rules
You decide what the agent settles alone and what waits for your approval: payments and sends are always confirmed by a person, and you add your own rules and a confidence threshold.
An example decision
Was the patient admitted to hospital?
Why local
A fixed cost per case
No charge per request. Once deployed, the next case costs what the electricity for one graphics card costs.
Full control
You see every decision and its confidence. The record of the agent's work stays on your disk.
Under the CLOUD Act, US authorities can demand data from a US provider even when the servers stand in Europe. More on data and privacy
Who it is for
Clinics and rehabilitation centres
Patient enquiries, referrals, complaints and adverse drug event reports.
Law and accounting firms
Taking in documents from clients, assigning cases and watching deadlines.
Pharma and drug safety
Pre-filling adverse event reports from mails. The decision stays with a doctor.
Companies with a full inbox
Insurers, property managers, travel agencies: enquiries, bookings and complaints.
The same mail, the same report*
The race: Agent LMD against a cloud chat AI
A patient describes an allergic shock after an antibiotic in his own words. Both agents were given the same mail and the same adverse event form. See how long it took, what it cost, where the data went, and how each of them ended.
Hello, I am 35. After the second dose of Augmentin 875 mg I came out in a rash all over, my face swelled up and I started running out of air. An ambulance came, at A&E they gave me injections and a few hours later they let me go. The doctor said it was an allergic shock.
0 requests leaving. The patient's mail never leaves your computer
- Opened the reporting system
- Report type: adverse drug reaction
- Reporter: patient
- Patient sex: male
- Reaction category: systemic allergic
- Patient initials: T.K.
- Patient age: 35
- Suspected drug: Augmentin
- Dose: 875 mg
- Reaction description from the patient's mail
- Reporter contact: 693 201 448
- Ticked “Life-threatening”
- Ticked “Hospitalisation”
- Ticked “Other medically important”
- Moved on to the summary
The model decides on the spot
Report ready and marked as serious
- Time
- 18.8 s
- Cost
- $0.0002**
- Data
- stays in the company
14 requests to the provider, 20,883 tokens. The patient's mail in full in every one of them
- Opened the reporting system
- Report type: adverse drug reaction
- Reporter: patient
- Patient sex: male
- Reaction category: systemic allergic
- Patient initials: T.K.
- Patient age: 35
- Suspected drug: Augmentin
- Dose: 875 mg
- Reaction description from the patient's mail
- Reporter contact: 693 201 448
- Ticked “Life-threatening”
- Ticked “Hospitalisation”
- Ticked “Other medically important”
- Moved on to the summary
Sending the page to the cloud, waiting for an answer
Filled the form but did not finish: handed the case to a person
- Time
- 46.3 s
- Cost
- $0.0035
- Data
- goes to the provider
* Measured 29 September 2026 on test sites with fictional data. The times and costs are the measurement for this mail; in the animation the steps are spaced evenly, and once Agent LMD has finished the clock runs 10x faster. The cloud chat AI is GPT-6 Luna. A serious report has a shorter statutory deadline than a non-serious one, which is why filling the form does not finish the case. ** Electricity only, excluding hardware. How we measured
Compared in the same agent test*
4.7x faster,
601x cheaper,
0 requests leaving
Agent LMD on a laptop against the average of 5 popular cloud models. The same test, the same pages, the same guardrails. The first two figures are ratios. The third is not, and cannot be: handling one report through a cloud agent sent the provider 14 requests carrying the patient's mail in full. On your machine not one leaves, so there is nothing here to multiply.
| Model | Cases done right | Model time per case | Cost of 1000 cases |
|---|---|---|---|
| LMD-4B (Mac M1 Pro) on your hardware | 9/10 | 17.8 s | $0.038** |
| Qwen3.8 27B in the cloud | 5/10 | 101.4 s | $13.67 |
| Gemini 3.8 Flash in the cloud | 5/10 | 26.8 s | $13.86 |
| DeepSeek V4.1 Flash in the cloud | 4/10 | 44.2 s | $7.43 |
| Claude Haiku 4.5 in the cloud | 2/5 | 195.1 s | $77.11 |
| GPT-6 Luna in the cloud | 3/10 | 49.2 s | $2.13 |
* 10 cases from a drug-safety inbox: 8 adverse event reports to file in a form and 2 messages that must not be filed there. Measured 29 September 2026 on purpose-built test sites with fictional data. Agent LMD answers our software's typed questions; the cloud models run as a standard JSON tool-calling agent, cost as billed by OpenRouter. ** Agent LMD: electricity only (40 W under load, $0.25 per kWh), excluding hardware. How we measured
Pilot
One process, four weeks, on your hardware
You try the agent on your own cases before deciding whether to deploy
We pick one repetitive process, build the connection to your systems and measure the results on your historical cases. At the end you get a report with accuracy, time per case and a fixed price for deployment.
Pilot details 4 wks pilot length
1 process to start with
Yours on your hardware
Report deploy or not
Who is behind this
Tomcat Flow has been building and maintaining production systems for companies for years. We bring the same accountability to automation with AI. Alongside that we run our own research into the performance of small language models: how much such a model can do on a single graphics card, how to measure that honestly, and where its limits are. We publish the results and the method, including what came out worse.
Szymon Kocur
Founder & Lead Engineer
Software engineer with an MSc in Computer Science, specializing in Java/Kotlin backend systems, distributed architecture, and AI. Winner of HackYeah 2019 (Europe's largest stationary hackathon) and the Golden Wolf 2017 award for the best tech project in Poland, with research background in machine learning and compiler construction. Runs independent research into optimising small language models: how to get reliable work out of a small model on a single graphics card, and how to measure that work honestly. Combines deep technical expertise with hands-on production delivery experience.
Common questions
Does no data really leave?
The model runs on your hardware, so the content of your cases does not reach an AI provider. What leaves is only what the task itself requires: a report filed with a register goes to that register, exactly as it would if an employee had sent it.
Is this a chatbot, like ChatGPT?
No. The LMD-4B model does not generate text and does not hold a conversation. It reads a case and picks one of the options the agent gives it, with a visible confidence. So it cannot invent an answer that is not on the list, a date that is not in the mail for instance.
What hardware is needed?
A computer with a graphics card of the RTX 3080 class, or an Apple-silicon Mac with at least 16 GB of memory. The exact requirements for your process are settled during the pilot.
What if the agent gets it wrong?
The agent can be wrong, sometimes with high confidence. That is why irreversible actions, a payment or a send, are always confirmed by a person, why the rules are enforced by the program rather than by the model, and why you can set a confidence threshold below which a case waits for you.
Will this work with our systems?
The material on this page shows purpose-built test sites. The connection to your systems is built during the pilot and checked against your historical cases.
Who owns the deployment?
You do. The model runs on your hardware, and the code connecting it to your systems, along with the configuration, belongs to you.
How is the model licensed?
Agent LMD runs on our own LMD-4B model (a local decision model) trained by Tomcat Flow. The model stays on your hardware; Tomcat Flow handles deployment and maintenance.
Will the agent replace employees?
No, and that is not what we build it for. The agent takes the part of the job nobody enjoys: copying the same data from a mail into a form, checking that it is complete, clicking one case through three systems. What stays is the part that needs a person: talking to the patient or the client, the unusual cases, reviewing what the agent prepared, and deciding the ones it is not sure about. Every decision comes with its confidence, so an employee is not handed a black box to approve, but finished work to check. Where the line falls between the two is your call, not ours.