AI Austin

Problem

A portfolio answers the questions its author anticipated. A conversation is different: visitors bring their own context and follow the thread that matters to them. AI Austin makes that exploration more natural while managing risks around impersonation, privacy, accuracy, and cost.

Insight

The useful product is a bounded interface to public evidence—not an autonomous copy of a person. It should know where its knowledge came from, admit uncertainty, and hand important questions back to Austin.

Decision

Start with text, add voice only after an explicit user action, and issue temporary session credentials through a small service. Keep permanent provider secrets out of the browser. Do not give the first version tools with write access.

Implementation

The deployed text experience includes a clear AI disclosure, explicit consent, streamed sentence-level replies, source and resource links, guardrail redirects, reset and withdrawal controls, and a visible direct-contact path.

Evidence

• A single Cloud Run service owns the interface and protected APIs.
• The public knowledge corpus is reviewed and attributable.
• The application does not store raw questions, answers, or transcripts.
• Automated evaluation covers grounded answers, no-source behavior, sensitive requests, latency, and response structure.

Reflection

The text experience is live and intentionally bounded. Voice remains a later phase requiring separate consent, provider evaluation, and explicit activation.