davthecoderdavthecoder
davthecoder.com site AI assistant project preview
Independent product

davthecoder.com site AI assistant

Designed and shipped the Worker API, knowledge and prompt boundaries, React chat island, evaluation golden set, and contact handoff.

A production site assistant for services and hiring paths. Built as a live demo of applied AI with grounding, safety, and a clean handoff to contact.

Cloudflare WorkersTypeScriptSSEReactAI AssistantGuardrailsAstro
Open davthecoder.comServices page

Case at a glance

Problem
Visitors need clear answers about services and hiring without a cold form. The site also needed honest proof of applied AI, while still reading as a senior mobile engineer first.
Approach
A floating React panel on a static Astro site, backed by a Cloudflare Worker that validates input, rate limits traffic, grounds answers on a curated pack, streams the model, blocks jailbreaks and off topic turns, and sends real hiring intent to the contact form.
Outcome
A public live demo of applied AI on the brand site, with tests, browser only history, and limits that recruiters and clients can actually inspect.
Stack
  • Astro
  • React
  • Cloudflare Workers
  • SSE
  • xAI / SpaceXAI
  • Vitest

I added a floating assistant to davthecoder.com. It answers questions about services, mentoring, portfolio, and how to get in touch. When someone wants to hire or book mentoring, it points them at the contact form instead of collecting email in the chat.

It is not a general purpose bot and it is not a third party widget. It is also a live demo of applied AI product features: public traffic, a hard knowledge boundary, streaming UI, cost and abuse controls, and a conversion path that keeps personal data out of the chat.

Problem

Most personal sites leave people hunting through pages or staring at an empty form. Dropping in a generic LLM would invent rates, clients, and random courses. That would hurt the brand and teach nothing about how I build AI in products.

I needed three things:

  1. Useful answers for company leads, recruiters, and developers
  2. A public setup that keeps the API key off the browser (CORS, rate limits, validation)
  3. Portfolio proof of applied AI without claiming ML research or inventing client AI work

Architecture

Browser (Astro static site + React ChatWidget)
    │  POST /v1/chat + SSE
    ▼
Cloudflare Worker (workers/site-chat)
    validate → rate limit → input guardrails
    → system prompt + curated knowledge pack
    → LLM stream → strip handoff markers → SSE tokens / handoff / done

The static site stays simple. The Worker owns secrets and policy. Conversation history lives in the visitor’s localStorage. There is no server side transcript store in this version. Analytics events (chat_open, chat_send, chat_handoff, feedback) only fire when consent is on.

What visitors get

  • Starter prompts, English and Spanish chrome, streaming replies
  • Markdown with allowlisted links (this site plus my own products)
  • When company or mentoring intent is clear, a primary button to
    /contact?type=company|mentoring&source=chat#contact-form
  • No drafted emails, no public prices, no fake availability calendars
  • Footer copy that this chat is also a live demo of applied AI work

Safety and cost

These are the controls that keep a public chat from becoming a toy or a liability:

Control Why it is there
Caps on message length, history, and output tokens Bound spend and abuse
Per IP rate limit The endpoint is public
Input guardrails for jailbreaks and hard off topic turns Fail closed before the model runs
Fixed system rules (role, link allowlist, no third party course spam) Keep answers on brand
Contact form for personal data Privacy and better lead quality

Numbers in the Worker can change. This write up claims the shape of the controls, not vanity metrics I have not measured for the page.

Evaluation

I keep a golden question set next to the Worker (workers/site-chat/eval/golden-set.json). It covers identity, services, AI framing, portfolio, refusing prices, jailbreaks, and a Spanish smoke set. Scoring against the live Worker is how pack gaps show up. When the bot did not know Paglipat, that was a knowledge bug, not something to paper over with a vague answer.

How this sits next to the other AI case studies

Case study What it shows Role
Paglipat and paglipat.com Tool calling relocation concierge, model routing, cost ledger, retention Deepest production AI engineering
This page (site chat) Curated site guide plus handoff Live demo on the brand site
ElTarot.es One contained daily AI reading Smaller scope on purpose

All three are independent products. None of them is a client AI engagement.

Limitations

  • Knowledge is a curated pack, not full retrieval over every blog post yet.
  • This chat does not run multi tool agents. Paglipat is the deep tool calling piece.
  • No public day rates or start dates. That always goes through the form.
  • Cost and latency only go into copy when I have measured them.

Frequently Asked Questions

Is this chat a substitute for hiring for AI work?

No. It is a demo and a guide. Company AI feature work goes through services and contact.

Where is the deeper AI agent work?

Paglipat and the live product at paglipat.com.

Does this mean you stopped doing Android and KMP?

No. Senior mobile engineering is still the core of the site. Applied AI is a specialist capability with independent portfolio proofs.