Carl Broker — carlbroker.com

I build AI systems that are meant to be used, not just demoed.

This is the source for my personal site — a working surface for the projects, writing, and experiments I'm shipping. If you're reading this on GitHub, the live version is the better way in:

→ https://cbroker1.github.io

About me

I'm an Advisory AI Engineer at IBM, based in Duluth, Minnesota. My work sits at the intersection of agentic AI, retrieval systems, LLM platforms, and production engineering — designing systems where models, tools, and data work together to solve real problems.

The path here ran through neuroscience research → healthcare analytics → applied research at the EPA → enterprise data science → enterprise AI. That mix gave me a strong bias toward systems that are grounded, testable, and useful in the real world.

These days I focus on things like:

• multi-agent orchestration
• retrieval-augmented generation
• embeddings and vector databases
• tool-calling architectures
• evaluation and guardrails
• production-ready LLM workflows

Where to find me

• Site — https://cbroker1.github.io
• GitHub — @cbroker1
• LinkedIn — Carl Broker

About this repo

The site is intentionally small and static — Astro, a handful of components, and Markdown for everything that's actually content. The design is "Hybrid Studio": a dark, atmospheric hero that hands off to a clean light editorial context for the writing and project detail pages. The whole thing is meant to read like opening the door to a working studio rather than a portfolio template.

Stack

• Astro — static site generator
• Markdown content collections for projects and writing
• Transformers.js — browser-local inference for the site assistant, loaded on demand and never on page load
• GitHub Pages — hosting

Local development

npm install
npm run dev              # dev server at http://localhost:4321
npm run build            # production build to ./dist
npm run preview          # preview the production build locally

npm run assistant:eval    # retrieval checks against recruiter-style questions
npm run assistant:test    # controller, grounding gate, and corpus provenance
npm run assistant:answers # print the assistant's deterministic output

Those three read dist/assistant/corpus.json, so run npm run build first.

There is also a browser-level check that drives the real panel — opening it, asking a question, resetting, and closing — over the Chrome DevTools Protocol:

npm run preview                                                   # note the port
google-chrome --headless=new --remote-debugging-port=9222 about:blank &
npm run assistant:drive -- 4321                                   # that port

Requires Node >=22.12.0.

Ask about Carl — the site assistant

A retrieval-grounded assistant lives in the bottom-right corner of every page. It answers questions about my work using only this site's own content, and it runs entirely in the visitor's browser — no backend, no API keys, no inference service, nothing logged.

How it works:

1. At build time, an Astro endpoint (src/pages/assistant/corpus.json.ts) collects the approved public sources, strips them to prose, chunks them, and emits /assistant/corpus.json.
2. In the browser, a question is scored against those chunks with a BM25F-style lexical retriever plus a curated synonym map. No embedding model is downloaded.
3. If the question's subject does not appear in the corpus at all, the assistant declines and says what is missing. No model runs.
4. Otherwise the top passages and the question go to a small instruction model running locally on WebGPU, which writes the prose — and only the prose.
5. Source links are rendered from retrieval metadata, never from the model, so a citation can never be invented.

Without WebGPU the assistant degrades to quoting the closest passage from the site. The rest of the site never depends on any of it: the assistant's code is dynamically imported and a failure anywhere in it is contained.

What it is allowed to know is defined in exactly one place — the header of src/pages/assistant/corpus.json.ts. To teach it something that has no natural home on a page, edit src/knowledge/curated-profile.md; see src/knowledge/README.md for the editing guide.

Project layout

src/
├── components/   global, section, card, and assistant components
├── content/      markdown for projects, writing, about, experience
├── knowledge/    hand-written public knowledge for the site assistant
├── layouts/      BaseLayout
├── lib/          assistant retrieval, prompting, and inference modules
├── pages/        routes (homepage, detail pages, assistant corpus endpoint)
└── styles/       design tokens and base styles

public/          static assets (fonts, video, favicon)
scripts/         offline checks for the assistant

Project and writing entries live in src/content/projects/ and src/content/writing/ as Markdown. Schemas are defined in src/content.config.ts.

Deployment

Static build deployed to GitHub Pages from master via GitHub Actions.

---

If you're building something that needs an AI engineer who cares about whether the system actually works — get in touch.