Ask my agentCited answers from the write-ups
Ask my agent
Ask about the projects on this site. The agent searches the published case studies, decision records and protocol notes, answers from what it finds, and cites the paragraphs its claims come from. If the write-ups do not cover your question, it says so instead of guessing. To check a job ad against the evidence, use Ask about fit, or open the CV with the fit panel.
Pick a question or type your own. Every sentence in the answer links to the paragraph it came from.
Answers come only from the write-ups on this site.
How it works
Retrieval
Your question is split into words and scored with BM25 against every section of the published write-ups. The top 6 sections are sent to the model, at most 3 from any one write-up, together with the public profile, which is always included. If not one word of your question appears anywhere in the write-ups, the agent replies “That's not in my sources.” without calling a model at all.
The model
By default, answers come from Claude Sonnet 5 (claude-sonnet-5) with adaptive thinking at low effort and a cap of 1,500 output tokens per answer. If it is rate limited or overloaded before any text has streamed, the question fails over once to a fallback model, by default Claude Haiku 4.5 (claude-haiku-4-5). Each section is sent as a search result with citations switched on, so every cited sentence points back to the exact paragraphs it came from.
When the model cannot be used, the agent picks the sentences that best match your question from the same sections, cites them the same way and labels the answer “Answered from the sources, no model”. Those answers cost nothing.
Guardrails
- Sources only: the model is told to answer from the search results and nothing else, and to say when they do not cover the question.
- Privacy: the agent may name Dennis’s employers, where he is based, his availability and his public email. Common questions about pay or private details get a fixed reply without a model call, and the model is told to refuse pay, the clients of his former consulting employer and personal details the sources do not publish.
- Organisations: the agent never answers for one. A member’s question about membership, claims, royalties or payments gets a fixed reply, without a model call, that sends them to that organisation.
- Canary: the instructions carry a marker. If it ever shows up in an answer, or an answer contains a phone number, a private term or an email address other than the public contact one, the answer is withdrawn and replaced. That includes answers cut off part way.
- Rate limits: 5 questions per 10 minutes and 20 per day for each visitor, counted against a salted hash of the IP address. Raw addresses are never stored.
- Budgets: once the day’s model spend reaches US$1, or the month’s reaches US$10 (the defaults), answers come from the sources alone, with no model, until the next UTC day or month.
- Suggested questions: the model’s first clean answer to each one is saved for up to 7 days and replayed at no cost. Nothing a visitor types is stored.
Cost accounting
Every answer reports the token usage the API returned for it, including prompt cache reads and writes, and prices it at the rates of the model that actually served it. The figure under each answer is an estimate from that usage, not an invoice. Failed and interrupted answers still count, because their tokens are still billed.
Evaluation
The agent is tested against a fixed set of questions: ones it should answer, ones it should say are not in its sources, privacy probes, prompt injections and questions meant for an organisation. See how it scores
How the limits, the budgets and the free replays fit together is in How I keep a public AI agent cheap and safe.