Capability
Applied AI
You want an assistant that answers from your own documentation. You also need to know what it will do when the answer is not in there, and to be able to prove to somebody that it is right often enough to deploy.
What we deliver
The work itself
- Assemble an evaluation set of real questions with agreed correct answers, before any model is chosen
- Build retrieval over your approved content, with chunking tuned against that set
- Report accuracy, refusal rate and citation coverage as numbers, and re-report them on every change
- Constrain the assistant to cite a source, or to say it does not know and hand to a person
- Log every exchange with its retrieved context, so a wrong answer can be traced to a cause
- Put a human in the loop wherever an answer commits the business to something
- Keep the content pipeline in your hands, so correcting a document corrects the assistant
- Re-run the evaluation set on model upgrades, so a provider change cannot silently degrade quality
How we work
What the engagement looks like
Phases, durations and what each one hands over are listed below. The stack is published in full — nothing about how this is built is a surprise once work starts.
Stack
What we build it with
- RuntimeTypeScript · Python · Claude API · Streaming responses over SSE
- DataPostgreSQL with pgvector · Hybrid retrieval (BM25 and dense) · Document ingestion pipeline · Conversation store
- PlatformCloudflare Workers · Azure Container Apps · Microsoft Graph for document sources
- OperationsEvaluation harness in CI · Prompt and retrieval versioning · Token and cost budgets per conversation · OpenTelemetry traces
Engagement shape
Phases and timeline
- Evaluation set1 to 2 weeksBetween 80 and 200 real questions with agreed answers, and a baseline score for doing nothing
- Retrieval build2 to 4 weeksIngestion pipeline, retrieval tuned against the evaluation set, and the measured score
- Assistant and guardrails2 to 4 weeksAssistant with citations, refusal behaviour, escalation rules and conversation logging
- Supervised pilot3 to 6 weeksLive with every answer reviewed, and a decision to widen or stop based on the review
What we need from you
- The documentation you are willing to stand behind as correct
- A subject-matter expert to settle disputed answers in the evaluation set
- A decision on what the assistant is never allowed to answer
- Someone to review pilot conversations daily while the pilot runs
After go-live
Who holds it at 3am
- The evaluation set runs on every prompt, retrieval or model change, and the score is reported
- Escalated conversations are reviewed weekly, and the gaps they expose become content fixes
- Cost per conversation is tracked against a budget agreed before launch
- If accuracy falls below the agreed floor, the assistant reverts to handing over to a person