No fabrication
Every quantified claim on this site traces to the CV or a project record. Where evidence does not exist, the claim is left out — the same rule binds the assistant.
Governance
How this site controls its claims
The Ask my CV assistant is a live sample of governed AI: grounded answers, deterministic controls, cited sources. This page states how it works, where it stops, and how crawlers may use the site.
The assistant answers strictly from the CV and case-study records published on this site. It is retrieval-grounded: it can only cite what the record contains, and it refuses what the record does not support.
Ask my CV
How an answer is produced
-
01 — Read a starter answer
The six suggested questions return a short excerpt directly from the same public records used by the site. They do not need AI generation. Other questions follow the steps below.
-
02 — Retrieve
Open-ended questions are matched to passages using semantic search, with keyword search as a fallback if the search service is unavailable. If nothing relevant is found, no answer is generated.
-
03 — Answer from the record only
The model receives the retrieved records and nothing else. It is instructed to answer strictly from them.
-
04 — Check the answer
The server checks the response format and that every cited passage was retrieved for this question. It rejects invented citation IDs, empty answers and refusals. These checks do not prove that every sentence follows from its sources.
-
05 — Cite
The answer links to the records it actually cites. Follow the evidence reveals the exact supporting passages and recorded project status, copied from the public records rather than generated by the model.
Cost and abuse controls: each visitor has a request rate limit, the assistant has a daily budget for generated answers. Suggested questions remain available when that model budget is exhausted. Repeat generated answers are cached against the question, current public record and answer policy. The crawler policy asks automated visitors to avoid the assistant endpoint; rate limits apply to requests.
Claim controls
The rules the whole site runs under
CV parity gate
A pre-release check compares selected site claims against the published CV PDF to catch inconsistencies.
Regression gates
Editorial, narrative and assistant test batteries run before every deployment, including grounding questions the assistant must answer and out-of-scope questions it must refuse.
Single source of truth
The assistant reads the same typed data records the pages render. The displayed source passages can be read on the linked pages; a pre-release check verifies that they are present.
Limitations
Where the assistant stops
- AI summaries can make mistakes. Where a summary and its source differ, use the source. Starter answers are direct excerpts.
- Search can miss relevant passages. Rephrasing a question may help. The evidence trail makes the selected passages visible; it is not independent verification of my work.
- It refuses anything outside the published record by design — including questions a conversation with me could answer.
- It answers one question at a time from the record; it does not hold a conversation history or learn from visitors.
- Open-ended answers use a hosted language model and wording can vary. Suggested questions use the public record directly.
This page records intent and known limits; it does not assert formal compliance with any standard. Anything the record cannot answer, I can — email me directly.
Crawler policy
How machines may read this site
Search-engine crawlers and AI search or user-fetch agents (services that fetch a page to answer a person and cite it) are welcome across the public site. AI training crawlers are currently permitted. No crawler may call the assistant endpoint: automated traffic must never consume the assistant’s budget.
The machine-readable version of this policy, listing each agent explicitly, is /robots.txt.