From deployment to transformation
Deployment made the meeting assistant and documentation agents available. Adoption is evidenced separately: 40+ people use the meeting knowledge assistant, while 30+ engineers use the documentation workflow. Transformation changed how product decisions were prepared and how technical documentation stayed aligned with code.
The design followed the decisions people already had to make. Source material came first, AI produced a draft or grounded answer, and a named person decided whether it was ready to use.
- Start with the constraint. The workflow shaped the intervention, not a preferred tool.
- Keep approval human. Product and engineering owners validate before action.
- Label uncertainty. Testing and targets stay separate from production evidence.
Architecture / technical evidence
Architecture diagram. The transformation map is a sanitised public reconstruction: transcripts ground the functional stream; repository changes ground the engineering stream. People approve every downstream action.
Sanitised decision records
- ADR-01 · Source before generation. Context: output could be untraceable. Decision: retrieval stays bounded to transcripts or repository context. Consequence: accepted output needs a source path; no public quality score.
- ADR-02 · Human approval. Context: output affects product work and code. Decision: Product and Engineering owners can reject or rework it. Consequence: accountability stays human and review cost remains.
- ADR-03 · Automation boundary. Context: Jira writes carry operational risk. Decision: entry stays manual; Jira MCP stays in testing. Consequence: slower entry, contained blast radius.
| Boundary | Evaluation criterion | Public status |
|---|---|---|
| Grounding | Answer or draft remains traceable to approved source context | Current / production |
| Approval | A named owner reviews before operational use | Current / production |
| Automation | No uncontrolled downstream write | Manual; MCP testing |
| Adoption | Reported use and scale stay separate from quality claims | Evidence with limits |
Interpretation. These are public controls and criteria, not executed test results. Raw evaluation data is not public.
Public artefacts. Review the technical case hub and the separate Voicebot-PY repository for a public grounded-retrieval implementation and test suite. Project source, raw evaluation data and internal ADRs are not public; this page does not imply otherwise.
Evidence, targets and limits
40+ people use the meeting knowledge assistant. Separately, AI-supported documentation expanded from three to 30 microservices with 30+ engineers using it. These are different evidence sets.
Scope: meeting workflow. Definition: reported users. Period / method: not public. Source: public project record; no client artefact is published. Contribution: workflow framing and controls. Limit: adoption count, not a business outcome.
Scope: pilot-to-production documentation scale. Definition: microservices using the workflow. Period / method: not public. Source: public project record. Contribution: implementation and validation model. Limit: scale does not prove quality or causation.
Scope: engineering adoption. Definition: reported users. Period / method: not public. Source: public project record; no client artefact is published. Contribution: workflow implementation and adoption controls. Limit: separate from meeting and backlog metrics.
+40%cycle time and sprint-capacity change in a separate AI-assisted backlog workflow
Scope: separate backlog workflow. Definition: cycle time and sprint capacity. Period / method: not public. Source: public project record; no client artefact is published. Contribution: grounded, human-validated workflow design. Limit: Pedro's effect is not isolated from other factors.
Client identity and internal context are excluded. The diagram is a public reconstruction. Both targets describe the measurement framework, not achieved outcomes.
Have a workflow that is not changing?
Start with the operating problem.
I can help diagnose the constraint, test whether AI fits and define the controls needed for adoption.