Notes

Building production AI agent infrastructure

What we've learned shipping Meridian — an MCP server with real auth, Stripe billing, RAG via Vectorize, and SSE streaming on Cloudflare's free tier. Plus side notes on photonic / quantum retrieval research.

2026-06-24 · 9 min read · quantum optics

The Quantum Split-Step Fourier Algorithm for Nonlinear Optical Waveguides

A technical briefing on arXiv:2606.24643: Biancalana's Quantum Split-Step Fourier algorithm propagates classical mean field + Bogoliubov matrices U, V of quantum fluctuations through the same split-step integrator used in nonlinear Schrodinger simulations, then reconstructs symplectic eigenvalues, Williamson modes, and von Neumann entropy for any spectral window. Lossless Gaussian regime, exact diagnostics, low effective dimension even for radiation spanning many Fourier bins.

2026-06-19 · 6 min read · quantum machine learning

Quantum Kernels are Spectral Tensor Networks

Bridging the Fourier and tensor-network views of quantum machine learning. We dissect arXiv:2606.20402, showing how entangling tensor kernels factorize Fourier coefficient tensors as Matrix Product Operators.

2026-06-01 · 8 min read · briefing

Quantum optics of chiral and antichiral waveguide arrays

A briefing on Wang, Hiltunen, and Schotland's arXiv:2605.29536 paper: single-photon scattering in arrays of one-way waveguides, where chiral geometry produces light-cone transport and antichiral geometry restores classical-optics-like interference.

2026-05-22 · 6 min read · briefing

One-photon communication in atomic media — Yale's first-principles fidelity bounds

A briefing of Hong & Schotland's (arXiv:2605.22797) first-principles QFT framework for single-photon propagation. They establish a universal fidelity formula across erasure and dephasing channels, and discover a fundamental fidelity floor of 1/2 in the strong coupling limit—critical bounds for the next generation of quantum routers.

2026-05-14 · 9 min read · new

100% on the Apple Neural Engine — eligibility is not placement

We followed Apple's four-constraint recipe to the letter, converted a transformer to MLPackage, and got 100% static residency on the first pass. Then we asked the CoreML runtime where it actually ran the ops. 0% on the ANE. Same recipe, scaled to ESM2-class size, scheduler flipped to 100% ANE placement. The post explains why eligibility and placement are different problems, and ships the Swift MLComputePlan tool that distinguishes them without opening Xcode.

2026-05-14 · 8 min read

Browser AI on a RAM diet — what we shipped in WebGPU, what we sent back to the server

We tried to run a 1.1 GB vision model in the user's browser. Most users have 16 GB of RAM and four other tabs. This is what worked (Mol* protein rendering, SmolVLM in VR), what broke (Moondream2 as the default, the Cache API on Safari), and the rule we use now: render on the client, infer on the server, in-browser inference only when latency is the constraint and the model fits in ~500 MB. With the OPFS shim and the commit that ripped the WebGPU inference stack.

2026-05-09 · 12 min read

Reading Loudon to improve a router — nine experiments, one ship

Read Rodney Loudon's Quantum Theory of Light end-to-end to mine ideas for photon-route + meridian. Ten experiments grounded in chapters 3, 5, 6, and 9. Eight came back negative or decorative — squeezing was decorative, photon-number metric strictly weaker than Gaussian fidelity, non-Gaussian heralded states didn't survive the eval scale. One — a g(1) coherence-time feature — shipped to meridian-orbital@3.1.0 with +108% discrimination over what it replaced. Honest accounting with every number.

2026-05-09 · 11 min read

AI agents vs. reCAPTCHA Enterprise — building a pharmacy MCP that ships

I tried to make Grok run my pharmacy errands end-to-end. The plan: search â′' cart â′' automated checkout, all chat-driven. The wall: VTEX's reCAPTCHA Enterprise enforces hostname binding server-side — your worker forwards a real captcha token, Google reports the origin doesn't match the site key, VTEX silently drops the email and returns 200 {}. The pivot: anonymous cart in-worker + a /checkout/cart/add URL hand-off to the user's browser. Ships ~95% of the value without a Browserbase-class headless browser.

2026-05-09 · 10 min read · new

AI for personal finance — Grok â′" Binance through Fly + Bright Data, locked behind a passkey

A single-tenant finance MCP that lets Grok read my Binance balance from a country Binance geo-blocks. Three layers, each solving exactly one problem: WebAuthn at the door (passkey-bootstrapped OAuth, single-user), Cloudflare Workers in the middle (auth + tools + KV), a tiny Fly proxy through Bright Data's static-residential AR IP at the edge so Binance's whitelist hears from one fixed IP forever. ~

6/month all in.

2026-05-07 · 10 min read

From 17% to 81% — calibrating an orbital skill classifier with simulations and online SGD

A skill-routing classifier had a planet bias — 17 of 18 panel skills got classified as planet. We tried two textbook physics frameworks (Vallado's CRTBP, Sears & Zemansky's spectral classification), neither beat heuristics. Retuned heuristics hit 50% on the panel, then 81% [95% Wilson CI 60%, 92%] on real labelled data. Then we wired an online-SGD loop on top so the classifier improves from every user click — with constant per-request cost and zero local training.

2026-05-07 · 5 min read

Meridian as a Grok connector — orbital skill routing inside Grok

Grok's custom MCP connectors take a server URL + an OAuth dance. We just wired Meridian's orbital skill router into it. The architecture, the OAuth values to paste, and the design choice that lets Grok users authorize with one click — no GitHub PAT, no jargon, no setup wizard.

2026-05-06 · 7 min read

GitHub Packages vs npm — should you publish to both?

We just shipped meridian-mcp 2.0.0 to both registries from a single tag push. The honest tradeoff: anonymous install still requires npm, but for everything else GitHub Packages has earned its slot — auth shared with your code, automatic provenance, free private packages, same registry as your Docker images. The full release workflow + the npm pkg set name rename trick that lets one tarball ship under two names.

2026-05-04 · 8 min read

Building a production MCP server on Cloudflare Workers (with auth + Stripe)

Most MCP tutorials show npx create-mcp and a Hello-world tool. Real servers in production need auth, per-key quotas, billing, and an upgrade path. Here's the architecture we use for ask-meridian.uk — entirely on Cloudflare's free tier — and the specific gotchas you'll hit.

2026-05-04 · 6 min read

LLM-authored skills + a classical orbital classifier — why we route this way

Most RAG systems either retrieve from a static corpus or end-to-end LLM-rank. Meridian does neither. Llama-3.3-70B authors fresh candidates per task; a deterministic JS classifier assigns each a celestial body class (planet/moon/trojan/asteroid/comet/irregular) based on physics-style features. Why the metaphor, why the deterministic step, and where it goes next.

2026-05-04 · 5 min read

Real RAG with Cloudflare Vectorize + AI Gateway in 80 lines

Workers AI for embeddings (free tier), Vectorize for semantic search (free tier), AI Gateway for caching + observability (free). The whole stack costs $0 until you cross 100k requests/month. Here's the wiring we use for meridian's RAG layer.