vibencode
GenAI implementation
Can we put what we already know to work?
“Make support smarter.”
A retrieval assistant over your ticket history and product docs, answering with citations, measured on a set of real tickets, and handing to a person the moment it is not sure.
The model is rarely the hard part. What decides whether an assistant is trusted is everything around it: which sources it is allowed to read, whether it can show where an answer came from, what it does when two documents disagree, and how it behaves at the edge of what it knows. A demo answers the questions somebody thought of. Production answers the ones nobody did, including the ones where the right answer is that it does not know. Most of the list below is that.
- 01Ground it in your data: retrieval over the sources that actually hold the answer, rather than a model asked to remember.
- 02Make it answerable for: citations on every claim, a confidence threshold, and a defined hand-off to a person below it.
- 03Build the evaluation before the interface, from real cases rather than invented ones.
- 04Wire it into the tools people already have open, so using it is not a separate decision.
- 05Ship behind a flag, watch it on real traffic, then widen.
- Solution architectureWhich sources are read, how they are retrieved, and where the answer stops and a person starts.
- Working implementationRunning where the work already happens, behind your access controls, on your real content.
- Evaluation & deployment planThe real cases it is scored on, the threshold it has to clear, and how you widen past the first team.
The first description of the problem is almost never the problem.
Frame and measure
Watch the work, define what good looks like, and build the evaluation set from real cases.
Build
Retrieval, grounding, citations, thresholds and the hand-off path, integrated where the work happens.
Prove and widen
Behind a flag with one team, tuned against real traffic, then opened up once it holds.
Hand over
Documentation, the evaluation set, and a team that can retune it without us.
That is the shape of a full engagement. If it is more than you want to commit to yet, the same people will spend two days on one piece of it first, and you are free to stop there. Start a 48-hour sprint
- Access to the real sourcesNot a sanitised export. The mess is the problem we are solving, and a clean sample hides it, including the doc last updated in 2019.
- Someone who does the work todayA few hours a week from a person who performs the task. They will tell you which answers would be unacceptable; a specification will not.
- A decision on the measureWhat number should move, and who agrees it moved. Settled at the start, because a project measured at the end is measured by the disappointed.
- A route to productionWhoever owns deployment, in the room early. Capabilities that cannot be shipped are expensive demos, and shipping usually changes the design.
What actually is GenAI implementation?
Putting a language model to work on something your organisation already does by hand, grounded in your own sources rather than in the model's memory, so every answer can point at where it came from. The implementation is everything around the model: retrieval, citations, a confidence threshold, a hand-off to a person, and a measurement that says whether it is good enough to trust.
How do you stop it making things up?
Ground every answer in retrieved sources, cite them, and set a confidence threshold below which it declines and hands to a person. Then measure how often it is wrong on real cases, publish that number, and treat it as the thing to improve rather than the thing to hide.
Can this run without our data leaving our environment?
Yes, and it is a design decision made at the start rather than a retrofit. Where residency, retention or key ownership are hard requirements, that constraint shapes the architecture from the first line. See Sovereign AI.
Which model do you use?
Whichever one measures best on your task, and we will show you the comparison rather than assert it. In most retrieval work the model is not the deciding variable: the quality of the sources and the retrieval step matter more. If the choice genuinely is the open question, that is applied research.
What happens when you leave?
You keep the evaluation set, the documentation and a team that has been doing the tuning alongside us for the last few weeks. If we have done this properly, the handover is uneventful.
We tried this last year and it did not stick. Why now?
Usually because it was measured on demos rather than on real work, or because it lived somewhere nobody goes. Both are fixable, and knowing what failed the first time makes the second attempt considerably cheaper. Bring the post-mortem to the first call.
If the thing needs to act (call tools, change records, run for minutes without a person watching), you are describing an agent, and the engineering is different in kind rather than in degree. That is agent orchestration.
Can we put what we already know to work?
Two days, one question, a written answer. No procurement.