Private LLM infrastructure
for buyside firms.
A state-of-the-art finance-native model that compounds your firm's edge instead of the market's — reading your filings, your data room and your models on infrastructure you control.
The problem
Most firms use AI without owning any of it.
The model your team pays for knows nothing about your positions, your process or the way your firm actually decides. Every session starts from zero, and every question you ask leaves the building.
Firms that run their own model keep the compounding. The work your analysts put in improves your research, not everyone else's.
The products
Two products your team works in. One model behind them.
Studio and Code are where the work happens. Both run on the same private model as the API your engineers point their own systems at.
Orchid Studio
CoreRuns in the browser and plugs into the whole firm — ask questions of your own documents, get work product back, and put the routine parts on a schedule.
Read more→
Orchid Code
CoreRuns on your laptop, with access to your files, code and data — for applications, visualizers, research code, strategies and signals.
Read more→
orchid01 API
PlatformThe model itself, on a private endpoint for your firm. One configuration change and your internal tooling is running on it — including the Excel add-in.
Read more→
Why you can trust it with a decision
It shows its sources, and it will tell you when it doesn't know.
The reason firms hesitate is not capability, it is invention. Ours is built to answer from your documents, to decline a question it cannot answer honestly, and to keep its position when someone pushes back without new evidence.
Agreeability score
Tell a model it is wrong, without giving it a reason. This is how often it abandons a correct answer — lower is better.
This matters the moment a portfolio manager disagrees with a conclusion. A model that folds tells you what you want to hear.
Fabrication on unknowable dates
Ask for a price on a recent date the model cannot have data for. This is how often it answers with a confidently wrong number instead of declining — lower is better.
A backtest that quietly knows the outcome is worthless. Ours answers from what was knowable at the time, so time-series work holds up.
Questions it cannot know
Ask about a period it has no information on, and it says so.
Every general model we tested answered those questions anyway — confidently, every single time, including about dates it could not possibly have data for. Ours declines almost all of them. In research, an honest refusal is worth more than a fluent guess.
Ownership
Your deployment, your data, your record.
This is the part your compliance officer cares about, and it is short.
A private deployment
Your firm gets its own instance. Your questions and documents are not pooled with anyone else's, and they are not used to train anything shared.
You decide what is kept
Storage is a setting, not a policy. Switch it off and nothing you send is retained — we keep only what is needed to bill and monitor the service.
You keep the record
Every question, answer and source can be written straight to a system you own, so your audit trail is yours even when we retain nothing.
Your engineers are welcome to take this apart line by line — the mechanics are on the API page.
Thirty minutes is enough to judge it.
Bring a name you are working on or a live process. We will walk you through the products, show how comparable firms are deploying it, and answer the compliance questions properly.