00, Company
A full stack software company specialising in AI agents
We build complete systems: backend, frontend, infrastructure, operations. And agents, up to autonomous ones, following approaches that have held up in production. One partner for the whole system, not just for the model.
00.1
Four things, not five
The shortest description of our work that still holds.
- Full stack
- Architecture, backend, frontend, data, infrastructure and operations from one team.
- Agents in production
- Evaluation, guardrails, observability and cost control from the start, not bolted on after the prototype.
- Sovereign operation
- LLMs run locally, in your own data centre, hosted in Germany or as a cloud API. Your data stays where you want it.
- Small to large
- From a single model inside an existing process to an autonomous agent system inside a product.
00.2
The difference between a demo and production
An agent that works in a pitch and an agent that still works after six months in production are not the same system. The difference is created in week one, not at the end.
Demo logic
- Success is demonstrated, not measured
- The prompt gets adjusted until the example works
- Tools without defined failure behaviour
- Cost per request unknown
- Decisions cannot be reconstructed afterwards
- No operating concept for the day after launch
Production logic
- Test set and automated evaluation from day one
- Context deliberately controlled instead of patched later
- Tools with defined failure and abort behaviour
- Cost per request as a metric in monitoring
- Full tracing of decisions and tool calls
- Versioning and regression tests on every model change
Every point in the left column can be addressed before the first line of code. Retrofitting each one costs a multiple.
00.3
Four stages, four entry points
The stages describe increasing autonomy, not a mandatory path. Each one makes sense on its own and reads as an entry point.
| Stage | What it is | Typical entry | Prerequisite |
|---|---|---|---|
| 01 Single model | A tightly scoped model for exactly one task in an existing process | Classification, extraction, text preparation | Sample data from the real process |
| 02 Agent on the web | Tool access in a web interface for a defined user group | Internal assistant, research across existing data | System access, role model |
| 03 Agent in the process | Carries out work steps independently, with approval points | Creating records, preparing replies for approval | A documented process, approvals settled |
| 04 Autonomous in a product | A scalable agent system as part of a product, multi tenant | Agent features inside a SaaS product, high throughput | Load profile, tenant model, operating concept |
00.4
Your data stays where you want it
We run our own GPU infrastructure and know these deployment models from daily operation. None of them is better in general. Which one fits follows from your constraints.
| Model | Data control | Operating effort | Typical case |
|---|---|---|---|
| Local, at your site | complete | with you | Data must not leave the building |
| On premise, your data centre | complete | shared | Integration into existing operations |
| Hosted in Germany | contractually clear | with us | GDPR clarity without owning hardware |
| Cloud or API | with the provider | minimal | Speed and model quality come first |
| Hybrid | by data class | shared | Sensitive work local, uncritical load external |
00.5
How we build agents
Ten steps, in this order. Starting at step three means paying a multiple for step two later.
- 01 Use case
- We start with the process and a measurable target, not with the model.
- 02 Evaluation
- The success measure and the test data are settled before implementation.
- 03 Context
- Deliberate control over what the agent sees at which point.
- 04 Tools
- Tightly scoped, robust, documented, with defined failure behaviour.
- 05 Scope of action
- Permissions, validation, abort conditions.
- 06 Approvals
- Human sign off where mistakes are expensive.
- 07 Observability
- Tracing of decisions, tool calls and cost.
- 08 Cost and latency
- Model choice per subtask, caching, smaller models where possible.
- 09 Operations
- Versioning, regression tests on model changes.
- 10 Scope of delivery
- A small usable stage first, then extend.
00.6
An agent at work, instead of a claim
A runtime panel shows an agent at work: tool calls with their latencies, one blocked access, a draft reply appearing character by character, and readouts for calls, cost and elapsed time running alongside. Before the first write the run stops and waits for an approval. You can decide yourself, but you do not have to.
00.7
Agents that get better in production
Every approval, every rejection and every correction becomes a test case. A candidate runs against the live version on the same set, and only what is better without becoming worse elsewhere is rolled out. The previous version stays available as a fallback. Without that loop a system does change, but nobody can say in which direction.
00.8
We run what we sell
The TimeTracking AI Agent connects to existing time tracking software, including legacy systems without a usable interface. People book their hours in a chat instead of a form. A product of our own in continuous operation is the one proof of operational capability that cannot be delivered as a slide.
00.9
Let us talk about the case, not the model
Thirty minutes in which we place your project. Including the answer that it does not need an agent, if that is the honest one.