Retrieval before generation
Grounded in your documents, your tickets, your schema — with citations the user can open. An answer with no source is a guess wearing a suit.
AI
Assistants and agents that are allowed to be wrong safely.
An assistant in production needs three things a demo does not: a bounded set of actions, a record of what it did, and a way to measure whether it is getting better or worse.
Softrear builds LLM assistants and agents for production use, with retrieval grounded in the client’s own data, explicit tool permissions, human approval on consequential actions, and an evaluation suite that runs in CI.
How it works
Bounded tool access
Every action the agent may take is declared, permissioned and logged. Consequential actions get a human in the loop by default, not as an upgrade.
Evaluation in CI
A graded test set that runs on every change, so a prompt edit that quietly degrades accuracy fails a build instead of a customer conversation.
Fallbacks that degrade honestly
When confidence is low the system says so and routes to a human. Confident wrongness is the failure mode that destroys trust in these systems.
Cost and latency budgets
Token spend and response time tracked per feature, with caching and smaller models used where they are indistinguishable to the user.
Guardrails
- Your data is not used to train third-party models.
- Prompt-injection testing on every retrieval surface.
- Full transcript and tool-call audit log, retained to your policy.
- A documented kill switch, tested before launch.
Also in AI
The data platform
Almost every stalled AI project we are called into has the same cause: the data exists, but nobody can say where it came from, when it was last correct, or who is allowed to see it.
AI readiness assessment
The cheapest way to find out whether an AI project is worth doing is to spend two weeks finding out, with people who will tell you if the answer is no.
Tell us what’s stuck.
A senior engineer reads every one of these. If we’re not the right fit we’ll say so, and point you at someone who is.
Start the conversation
No spam, no drip sequence. A senior engineer replies within one business day.