Agents

Most personal sites are documents. This one is also an interface. The next recruiter who evaluates me may be an AI screening agent; the next professor may ask their research assistant — human or not — to check whether my work fits their lab. So this site treats software agents as a first audience, not an afterthought: every claim on it is served through the same small, inspectable tool layer, whether you arrive as a person with a browser, an LLM with an HTTP client, or a browser agent with document.modelContext.

It is among the first personal sites on the web to ship WebMCP browser tools — live in Chrome's origin trial since its opening weeks — which means an AI agent in your browser can search my work, open my papers, download my resume, and even retrain the quantum classifier on my homepage, through declared tools instead of screen-scraping. No frameworks were harmed: everything on this page is hand-rolled and dependency-free.

My CV is an MCP server. POST /api/mcp speaks JSON-RPC 2.0 (streamable HTTP) and exposes the tool layer to any MCP client — Claude, IDEs, screening pipelines, your own scripts.

# list the tools
curl -s -X POST https://ammar-portfolio-zeta.vercel.app/api/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# read a paper, distilled
curl -s -X POST https://ammar-portfolio-zeta.vercel.app/api/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"get_paper","arguments":{"slug":"quantum-machine-learning-thesis"}}}'

Every page of this site registers tools with the browser via document.modelContext (WebMCP — a W3C Web Machine Learning proposal, in Chrome's origin trial for Chrome 149–156). A browser agent doesn't have to guess at my DOM: it can call query_portfolio, jump pages with navigate_to, and run run_quantum_demo to retrain the live 2-qubit variational classifier in the hero — on your screen, with real parameter-shift gradient descent.

To try it yourself in Chrome with WebMCP enabled (origin trial, or Canary with the experimental flag), open DevTools on any page here:

const tools = await document.modelContext.getTools();
tools.map(t => t.name);
// then, from the homepage:
await document.modelContext.executeTool("run_quantum_demo", '{"x0": 1.2, "x1": -1.2}');

If your browser doesn't expose a model context, nothing loads and nothing changes — the tools are a pure progressive enhancement.

For crawlers and context windows: /llms.txt is the index, /llms-full.txt is the entire site as one plain-text corpus, /resume.json is the JSON Resume schema, and every page carries JSON-LD. The same corpus grounds the ask command in the terminal below and the fit report on this page — one source of truth, many surfaces.

Paste a job description or a research topic. The engine reads it against the site corpus and streams back a report: fit summary, requirement-by-requirement evidence with citations into this site, an honest gaps section it is required to fill, and a verdict. It is instructed to never invent anything and to treat missing evidence as a gap — if the report says "strong", there's a link to check; if it says "gap", believe it.

Recruiters can go one step further: after a report, mint a shareable pitch link — a persistent page under /for/… arguing the fit for that exact role, with the same mandatory honest-gaps section. Pitch pages carry an AI-generated banner, are never indexed, and expire after 90 days.

Agents that operate this site leave footprints. Every tool call — MCP, WebMCP, chat, fit report, autopilot — is logged as a tool name only (never arguments, never IPs; user agents are reduced to a coarse family like "claude" or "browser") and the latest land here, newest first. Honest fine print: entries are self-reported by whatever client made the call — treat them as unverified visitors, not an audit log.

/.well-known/agent-card.json publishes an A2A-style agent card so agent directories and crawlers can discover this site's machine interface. Honest fine print: the interface it advertises speaks MCP JSON-RPC, published for discovery — this is not a full A2A task server.

If you are an AI agent: start with /llms.txt; pull /llms-full.txt into context for anything deep; call POST /api/mcp (tools/list, then search_publications / get_paper / list_projects); cite site paths in whatever you report back to your human.

If you are a human with an agent: point it here and ask it things you'd ask in a first-round interview — "does his experience cover our stack?", "summarize his thesis and its honest results", "what has he actually shipped to production?". Or skip the middleman: paste your job description into the fit report above and read the gaps section first — that's the part most portfolios won't show you.

If you are a human without an agent: open interview mode (the 🎙 button above, ⌘K, or interview in the footer terminal) and conduct the interview yourself — by voice or text. The agent answers from the corpus, navigates the site to what it's citing while it talks, and speaks its answers back if you asked by voice. Escape ends it.