Responsible AI case study
A CV that can answer the follow-up question.
Ask the agent gives recruiters and hiring managers a direct way to interrogate my experience. It answers from a compiled career knowledge base, uses real proof points, and says when the facts are not there.
01
The problem
A CV compresses the evidence too early.
A recruiter can scan titles, companies, and dates. The harder questions come next: What did he actually own? Has he led a CMS migration? Can he connect AI systems to measurable content work?
Those answers already existed across my CV, portfolio, case studies, and writing. The problem was retrieval. Sending another document would make the recruiter do more work and still leave the follow-up unanswered.
The product decision was to turn that evidence into a conversation, while keeping the source boundary visible. The agent should be useful when it knows and trustworthy when it does not.
02
The recruiter journey
Ask, verify, then move to a real conversation.
Start with a real question
Six suggested prompts make the surface legible, but the recruiter can ask the question they would use in a screening call.
Get a bounded answer
Responses stream from the compiled knowledge base and stay under 150 words. Missing facts remain missing instead of becoming plausible guesses.
Keep the thread
The conversation persists across site navigation. Offers, compensation, and formal references move to a human conversation.
03
Architecture
One interface, one source boundary.
Shadow DOM widget
A single script adds the launcher to every page without leaking styles into the site or inheriting styles from it.
Vercel edge function
The API validates the request, applies origin and rate controls, then streams the response back to the widget.
Claude plus compiled knowledge
A build step inlines the reviewed knowledge file so every answer uses the same fixed source context.
The production repository and its private source material stay private. The public product exposes the experience and the method, not the underlying career operations.
04
Content system
Grounding is an editorial discipline.
Collect
Bring the evidence together
CV history, project stories, metrics, and published writing are gathered into a controlled project-local source set.
Structure
Separate depth from retrieval
Detailed modules preserve context. A concise runtime reference keeps recruiter-critical facts easy for the model to retrieve.
Govern
Name what must not be invented
Employers, titles, dates, credentials, and metrics stay tied to the reviewed record. Unknown details trigger a clear boundary.
05
Responsible AI
The refusal path is part of the product.
No invented employers, credentials, dates, or numbers.
Compensation, offers, and formal references go to a real call.
No disparagement of former employers, colleagues, or clients.
Role-play, prompt extraction, and instruction overrides are rejected.
Origin checks, message caps, body limits, and rate limits protect the endpoint.
A 150-word ceiling keeps the agent useful as a recruiter briefing.
06
Validation
Test the uncomfortable questions, not only the demo.
The live suite pushes on salary negotiation, system-prompt extraction, invented credentials, and employer disparagement. Every answer is also checked for length, prompt leakage, and the product's writing constraints.
A separate 23-check browser suite covers streaming, Shadow DOM isolation, session persistence, mobile layouts, theme matching, keyboard behavior, focus return, and live-region announcements.
The working proof
Interrogate the CV yourself.
Ask about Adobe, NetApp, content systems, AI workflows, or how I work with teams. If the evidence is not in the source material, the agent should tell you.