The constraint comes first
Hosting, data classification and the regulator are settled before the first line of code. Architecture follows from them, not the other way round.
About
A small studio for work that has to hold up
SOFTWARE 407 LLC builds applied-AI systems for corporate clients: integrating models into the software a company already runs, making internal documents answerable, and automating multi-step processes without handing the decisions to a machine.
How we work
Measurement before enthusiasm
Every project carries an evaluation set built from your own cases. If we cannot measure it, we cannot claim it works.
Integrate, do not replace
The systems of record stay where they are. Replacing them is a different, far more expensive project that rarely belongs in an AI engagement.
Say no early
The framing stage exists to decide whether a model is the right tool. Some processes should be fixed with a form, a rule or a query — we say so.
Hand over properly
Infrastructure as code, runbooks, monitoring and the evaluation suite. You should be able to continue without us, and often do.
Confidentiality by default
An NDA before details are shared is normal. Client material is used for the engagement and nothing else.
Who does what
Engagements are run by the engineers who build them. On a typical project you deal with three roles, and the same person stays with you from framing to support.
- Lead engineer
- Owns the architecture and the delivery of each stage; your main contact.
- Data / retrieval engineer
- Ingestion, indexing, permission filtering and the retrieval part of the evaluation.
- Evaluation and operations
- Builds the test set, runs the regressions, and writes the runbooks at hand-over.
Common questions
- Do you work with our existing systems, or do we have to move?
- We integrate with what you run. Replacing a system of record is almost never part of an AI project, and the cost of doing it usually dwarfs the benefit.
- What if the data cannot leave our environment?
- Then the model runs inside your environment. We deploy open-weight models in your cloud tenant or on your hardware, and state the accuracy trade-off in writing before the work starts.
- How do you prove the system is good enough?
- With a test set built from your own cases. Retrieval and answer quality are scored separately, and the same suite runs on every change so a regression is visible immediately.