We design and build custom software around large language models for organizations that cannot afford a black box. The systems we ship do real work, show where every answer came from, and keep client data on hardware our clients own.
Two fields where the work is repetitive enough to automate, and consequential enough that nobody will accept an unverifiable answer.
Litigation and insurance-defense practices carry more document work than the hours in a day allow. We build workspaces that produce the first draft of the motion, the report to the carrier, the correspondence, with every sentence traceable to the record it came from.
Schools face the same shape of problem under stricter rules about children's data. We build tools that return hours to teachers while keeping the professional judgment, and the accountability, with the educator.
Three commitments that shape the architecture, not the marketing copy.
The rules of the field are design inputs from the first day, not a review at the end. That means verification gates before anything leaves the building, records built to be handed to an auditor, a carrier, or a court, and documentation of what a system does and does not do. We will complete a client's vendor security questionnaire and talk directly to their counsel, IT reviewer, or insurer.
Client files stay on hardware the client owns. We do not host client records, and we do not create a cloud copy of them. Access is least-privilege, connections to systems of record are read-only by default, anything that writes is gated behind explicit human approval, and actions are logged.
The useful work is in the last mile: the specific workflow, the specific document, the specific person's day. We build against a real process, deploy in days rather than quarters, and measure whether the hours actually came back. There is nothing artificial about the result. The intelligence in these systems is the client's own process and judgment, implemented carefully enough to run at full strength.
Our deployments run on machines the client owns. The application, the case or student files, local transcription, and the working database all stay on that hardware.
Model calls go to Amazon Bedrock, running Anthropic's Claude models under enterprise terms that bar the use of customer prompts for model training. Amazon Bedrock holds SOC 2 Type II and ISO 27001 certification, and ISO/IEC 42001 certification for AI management systems. Those audit reports are Amazon's, and we provide them on request alongside our own written description of exactly what data transits and what never leaves the building.
In plain terms: prompts transit to Amazon under no-training terms. The underlying files never leave the client's machine. We put that in writing, with the contract provision cited, before a client asks.
We would rather hand a client a system that shows its work than one that asks to be trusted.
A native workspace built for a New York insurance-defense practice. Per-matter AI sessions that see only their own matter, drafting with provenance-typed citations, an attorney sign-off gate before anything counts as final, local dictation, docket monitoring, and per-matter economics. Runs entirely on the firm's own machine.
Our own product for schools. AI-assisted grading where the teacher owns the rubric and reviews every result, built around student-data safeguards.
Eric Ritter, founder
[email protected]
631-461-1838
East Setauket, New York
Engagements are scoped and priced per project, with support arrangements after delivery. We are glad to walk a prospective client's technical or compliance reviewer through the architecture before any commitment.