These are not videos or screenshots. Each one is a deployed agent running behind a rate-limited API, doing the work in front of you. Type something of your own and watch what comes back.
The AI systems shown here were designed and built by Christopher Grant, GORGES Chief AI Officer.The first is live and running now, and you can type your own pricing model into it. The second reads the same engine in reverse to audit what was actually charged. Both are live, and you can type your own into either. Underneath both: the model reads the language, and code computes every dollar.
Describe a pricing model in a sentence. Get a structured model and a correct sample invoice.
| Factor | Say it like this | What the calculator does with it |
|---|---|---|
| Flat recurring fee | ‘charge $500 a month’ | Adds it to the invoice before any usage is counted. |
| Per-unit usage | ‘$0.002 per API call’ | Multiplies the billable units by that rate. |
| Volume tiers | ‘$0.002 up to 1,000,000, then $0.0015’ | Prices each band on its own and adds them up, rather than repricing everything at the cheaper rate. |
| Tier discounts | ‘10% off calls above 1,000,000’ | Takes the percentage off that band alone, and shows it as its own line. |
| Free allowance | ‘the first 5 seats are free’ | Subtracts them before charging begins. |
| Monthly minimum | ‘with a $500 monthly minimum’ | Trues the invoice up if everything else falls short of it. |
Combine as many as you like in one sentence. The unit you name sets the example usage the invoice is priced against, so seats are billed against 25 seats and API calls against 1,500,000, rather than one figure being forced onto both.
The API key sits in the agent, never in this page. Input is capped and rate limited, and the text you type is treated as data to read, never as instructions to follow.
For production we would add real per-user authentication, a secret rotation cadence, full request logging, and a data-handling policy if real customer data ever flowed. This demo runs on fabricated data only.
The same pricing engine, run in reverse. Give it the rule these accounts were billed under and what each one was actually charged, and it recomputes every row and flags the money slipping through.
Every figure on this card is computed in code and reconciled: charged plus difference equals expected, on every row. The model writes one sentence at the end, from numbers it was handed, and it is told never to calculate one of its own.
For production we would add real per-user authentication, a full audit trail of every run, and a data-handling policy before any real billing data went near it. This demo runs on fabricated accounts.
A language model is very good at reading a sentence and very bad at being a calculator, so we let it do only the first job. It turns your words into a structured model, that structure is validated against a fixed schema, and then ordinary deterministic code computes every figure. Any total the model volunteers is thrown away. That is why the two demos above always agree with each other, and why you can audit every number they show.
Ask a question and get an answer grounded only in Daniela's published work, with its sources shown.
This is the public slice of Ask Daniela, the real assistant we run on the Great Energy site, one of our own companies. The public demo indexes only already-published material, in its own isolated index, with no connection to the private one. The full assistant runs on her complete body of work, privately.
What is deliberately not here: class transcripts, personal notes, student names, and anything unpublished. If it is not already public, it is not in this index.
For production we would add real per-user authentication, a secret rotation cadence, full request logging, and a data-handling policy if real client material ever flowed.
This is the content desk from GLIV, the platform we run a side project on. Pick a post type and a voice, choose what it may draw on, and it drafts a post in that voice, ready for review. It delivers nowhere.
Typography is enforced in code rather than asked for in the prompt. Em dashes and curly quotes are stripped after the model writes, because a model emits them whatever the instructions say, and an em dash is this brand's number one auto-fail.
What we left out on purpose: scheduling, posting, the customer database, the funnel, and the morning brief. Those touch real accounts and real data, so they stay a case study rather than a public toy.
The whole thing is one Cloudflare Worker in front of a static site, so the page and the API share an origin and there is no CORS dance and no separate backend to keep alive. The Worker holds every secret; the browser holds none.
The library. Our published articles, video transcripts and story bank were chunked at about 1,100 characters with overlap, embedded with bge-large-en-v1.5, and written to a Vectorize index with the text carried in the metadata so the writing model can actually read what was retrieved. Each chunk is tagged with a content type, which is what the 'Pull from' checkboxes filter on. No single page may contribute more than 12 chunks, because one page rendering a whole food database would otherwise be most of the index.
The retrieval. Your topic is embedded and matched by cosine similarity. Ask Daniela, the demo above, uses a grounding floor of 0.62: measured on this index, real questions score 0.68 to 0.92 and an off-topic question scores about 0.45, so below the floor it declines to answer rather than inventing something.
The rule that matters. A prompt asks, code guarantees. Anything that must be true is enforced after the model, not requested before it: em dashes and curly quotes stripped, links removed, hashtags removed, paragraph breaks inserted, invented people scrubbed, and the review status overwritten with a constant. That last one is the point of the demo. When a model returned a review block claiming it had published to Instagram, the code threw the claim away.
The stack, in full. Cloudflare Workers, Workers Static Assets, Workers KV, Vectorize, Workers AI (bge-large-en-v1.5 for embeddings, Llama 3.3 70B for the content desk, gpt-oss-120b for Ask Daniela), Europe PMC and Wikipedia for research, the Pexels API for photos, and the HTML canvas for the graphic. n8n runs the same agents as deployed workflows, and the Worker prefers them when their webhooks are configured.