PAI is the chat interface to this studio site. Not a chatbot bolted onto pages: the conversation is the navigation. Ask about the work, ask about methodology, ask what’s relevant to your context. The system retrieves grounded answers and opens the work in a reading pane beside the conversation, routing to what you actually need rather than asking you to find it.
Conversation as navigation
Questions replace browse literacy: the visitor doesn’t need to know where the work section lives.
Show, don’t tell
Six panel types surface real portfolio pages, kits, and generated HTML mid-thread, not prose descriptions of links.
Keyword corpus at scale
About 5,000 curated entries, intent modes, always-include pinning, and live annotations: retrieval design, not embeddings by default.
Voice that reads as speech
Talk or type via Deepgram (STT + Aura TTS), plus optional live voice over LiveKit: normalization for pace and pause, not flat text-to-speech.
01: The Problem
Static portfolios present. PAI routes.
A traditional portfolio shows the same content to every visitor in the same sequence. A hiring manager evaluating fit for a principal design role and a practitioner researching validation methodology will both start at the same homepage, navigate the same work grid, and either find what they need or not.
This forces a particular literacy. Visitors have to know to look for the work section, understand what a design kit is, and read an About page in a particular order. Most don’t have that patience. The portfolio was organized around what the designer wanted to show, not what the visitor needed to find.
The format also flattens context. A recruiter and a practitioner need to hear different things about the same project. Scope of ownership, decision-making authority, how engineering was involved: these answers change depending on who is asking and why. Static content can’t make that distinction.
02: How it Works
The panel opens before the AI references it.
A recruiter asking about healthcare AI work gets different material than a practitioner asking about validation methodology: same corpus, different retrieval path, different framing. The routing intelligence is the product.
The panel system extends this. The AI doesn’t just talk about work: it opens it. A piece of work referenced in conversation opens in a reading pane the visitor can read and scroll independently while the conversation continues. The AI knows what’s visible, scrolls the panel to the section being discussed, and respects when the visitor dismisses the panel. The conversation and the content are genuinely coupled.
Chat and summaries stay public; full write-ups and design-kit depth open after portfolio access: the same gate the site uses elsewhere, surfaced in-panel when you ask.
PAI is a retrieval-augmented portfolio chat. A question pulls the strongest chunks from a keyword corpus of about 5,000 entries, Claude streams the reply into the bubble, and when the answer needs proof the system opens the work in the reading pane: then scrolls to the section under discussion.
Talk or type both work: Deepgram handles push-to-talk transcription and Aura playback; optional live voice runs over LiveKit when Hear PAI speak is available. WordPress remains the source of truth: panel pages are proxied, chrome stripped, typography cleaned.
Ask, retrieve, stream
The turn hits a single streaming endpoint. The backend scores the corpus, builds a grounded prompt, and Claude replies token-by-token over SSE. Replies stay short by default so voice stays listenable and hiring managers aren’t stuck in a monologue.
Panels mid-reply
When the answer needs the artifact, Claude opens it in the panel before talking about it: write-ups, kits, diagrams, contact. Section highlights keep the scroll in sync with the prose. If Claude is unavailable, OpenAI continues the text stream without panel tools.
Recap after unlock
After portfolio access, Save a recap (or an in-chat shareable-recap ask) builds a private timeline post from visit signals plus the chat thread, not Claude, not a full transcript. Email transcript is a separate opt-in for the log alone.
03: What it Demonstrates
The portfolio demonstrates the practice it describes.
There’s a gap in most design portfolios between “I designed AI features” and “I understand how AI systems actually work.” PAI closes that gap by being a system built and shipped solo, not a feature designed inside a product org. The architecture decisions (SSE streaming, keyword retrieval over embeddings, tool calling for panel management, the voice normalization pipeline) are direct expressions of product judgment about this specific use case.
For principal-level roles at 0→1 AI companies, where the job often involves making these architectural calls before a team is in place: that’s the relevant kind of proof. The write-up and the product are the same artifact. Reading about how PAI was built while talking to the system it describes collapses the distance between explanation and evidence.
Short replies by default
The primary visitors are time-boxed. A response that goes long signals the AI doesn’t know when to stop. The model is capped near ~150 output tokens so answers stay short, offer to go deeper on request, and stay listenable in voice without a separate summarization step.
Keyword retrieval over embeddings
Keyword scoring with intent boosts gives direct control over what surfaces for which questions: a recruiter asking about “AI work” reliably surfaces different chunks than a practitioner asking about “validation methodology.” The tradeoff is manual annotation; the benefit is predictable, auditable behavior.
Fast opening greeting
The first message is assembled in code: time of day, device, unlock state, light geographic context, with no LLM call on open. First-token latency on a portfolio page is disproportionately impactful. A near-instant greeting is worth more than a slightly more tailored one that waits on model retrieval. Deeper personalization waits until the visitor actually asks.
Ask it something.
PAI has full context on the work: projects, methodology, background, and the build itself. Ask about the healthcare AI work, ask how validation methodology was applied on a specific project, ask what the studio is working on now.