How the AI works
Why a model at all
A rules engine can find the word “churn”. It cannot find a churn problem phrased as “logo retention normalised across the cohort”, and it cannot tell you whether a monthly churn rate of 3.1% is high for a vertical SaaS at seed, because that judgement lives in the thousands of decks the model has seen and in the framework of what a healthy SaaS retention curve looks like at this stage.
The second half is harder. Explaining that a single customer at 22% of ARR means the growth story resets if that customer leaves at renewal, in three sentences, with the arithmetic done, is a writing problem. That is the part investors pay for and it is the part only a language model does.
So the model is the engine. Everything else in this product is the work around it: getting the text out of a PDF, keeping a fifty-page data room in one request, and refusing to show you a quote that is not in your document.
The four steps
The whole document goes to a balanced model in one pass. The system prompt names what the memo must include: a plain-terms business summary, the market and comparables, the key metrics pulled from the text, the traction read in context, and the risks. The model is told to quote verbatim from the source and to leave out anything it cannot quote.
A separate pass focused on risks: what the deck says, what it implies, and what it does not say that a diligent investor would expect. Each risk is returned with a severity, a detail paragraph and an evidence quote. Evidence that does not appear in the source document is dropped before you see it.
Two to four sentences, in plain English, saying what the company does, who it sells to, and why now. This is a separate step because the business summary and the full memo have different requirements for length and register, and sharing a prompt would compromise both.
Cheap triage: before a memo is built, a fast model reads the first part of the deck and returns the sector and stage. This costs a fraction of the memo step and lets the routing layer calibrate the system prompt and the comparables framework for the right kind of deal.
Routing
This is not a diagram somebody drew. It is rendered from the same routing table the memo engine reads, so if it is wrong here it is wrong in production. Changing a route is one line in the provider file.
| Step | Model | Provider | Tier | USD per M tokens |
|---|---|---|---|---|
| Read the deck and data room and structure the memo | gpt-4.1-mini | openai | balanced | $1.60 |
| Surface the risks and red flags | gpt-4.1-mini | openai | balanced | $1.60 |
| Write the plain-terms business summary | gpt-4.1-mini | openai | balanced | $1.60 |
| Work out the sector and stage | gpt-4.1-nano | openai | fast | $0.40 |
Above roughly 30,000 characters, about sixty pages, the memo step escalates to the frontier model rather than the balanced one, because long-document accuracy matters more than cost at that length. Memos are capped at 60,000 characters in one pass, and anything beyond that is reported rather than dropped silently.
If the routed model fails or returns nothing, the call walks a chain of candidates from other tiers and other labs before giving up. One provider having a bad afternoon should not cost you your memo, and it does not.
Every evidence quote is located in your own source document before it is shown, and the memo structure is validated against a schema. A quote that does not appear is dropped and counted. This is the only part of the pipeline the model does not get a vote on.
Candidates
Every model below sits behind the same interface. Adding a lab is one case in one file, which is the entire point of building it this way. Meridian is model-agnostic on purpose, so a lab changing price or policy is a configuration change, not a rebuild.
openai · 1,000,000 token context
openai · 1,000,000 token context
openai · 1,000,000 token context
anthropic · 200,000 token context
google · 1,000,000 token context
meta · 128,000 token context
In this deployment the OpenAI and Anthropic adapters are written and the OpenAI one is live. The rest are declared with their real model names and switch on when their key is present. No model is trained on customer decks or data-room documents, by us or by our providers.