Decoy sits between you and frontier models. Sensitive details are swapped for opaque tokens before they leave your device, so the model can do its work without ever seeing what's underneath.
The Moment
It's 3pm. The board meets at 9am tomorrow. You need a Q4 forecast across your top accounts and a read on which one would hurt most if it churned. The data's in a spreadsheet. The names are real. The numbers are real. The fastest path to a useful answer is ChatGPT.
You hover over paste. You delete the column with the names. You paste anyway. You delete the chat afterward, like that helps. Everyone you know does some version of this. None of you talk about it.
What's Broken
Frontier AI is the most capable tool that's ever sat on your laptop. The catch is the only data it's actually useful for is the data you shouldn't be sharing with it. The customer list. The salary review spreadsheet. The board pack. The patient cohort. The contract terms. The deal pipeline.
Your two options today are bad. Paste it and hope. Or don't use AI and lose the productivity. Most people pick option one and feel weird about it. Their compliance team isn't telling them that's fine.
The Idea
Decoy lets the cloud LLM do the thinking on opaque tokens. Your machine holds the real values and runs the maths.
Watch It Work
A worked example using a synthetic but realistic sales spreadsheet. The whole thing in 60 seconds.
On this question, the cloud's reasoning was wrong, it picked the wrong account as highest-risk. Decoy caught it because it was running the maths against ground truth, not against a guess. Privacy was the goal. Better answers were the surprise.
What This Does
The cloud model never sees a name, a number, a customer, or a salary. The mapping is held in your browser memory and destroyed when you end the session. It's not "we promise we won't look." It's "there is nothing for anyone to look at."
LLMs are unreliable at arithmetic. Decoy doesn't ask them to do it. The cloud plans the analysis. Your machine computes it. Every formula is visible, every input is auditable. You can show your CFO the working.
You stop deleting the customer column. You stop hesitating before paste. You stop deleting chats afterward. The architecture quietly does the work the policy was supposed to.
What It's For
Forecasting, comparison, ranking, aggregation, variance analysis, cohort summary, simple statistical work. The kind of question you'd otherwise put in front of Excel, but you'd rather ask the way you'd ask a smart colleague over Slack.
Decoy isn't for free-text generation. It isn't for code review. It isn't for conversational AI. It's for the moment you have a spreadsheet, a question, and twenty minutes before someone needs an answer.
Where It Fits
Why Trust Decoy
"Trust us" is the weakest possible promise. Decoy doesn't make it.
The mapping between real values and tokens is held in your browser memory and destroyed on session end. There is no Decoy server in the data path. The tokenisation engine is open and inspectable. Network logs show exactly what's been sent to the cloud and what hasn't. You bring your own API key, your chosen LLM provider's terms apply, and Decoy doesn't intermediate.
The architecture is the trust signal. The product would not work if the architecture broke.
Early Access
Decoy is in private alpha. We're letting people in slowly because each early customer shapes what gets built next. We're looking for finance leads, operators, and analysts working with sensitive data on a daily basis. People who'd actually use this and tell us where it's wrong.
If that's you, send a note. Tell us what you do, what you'd reach for it on, and what you'd want it to do that it probably doesn't yet. We read everything. We reply fast.
Built in Sydney. Working in public.