Glyphh AI Launches Ada, a Classifier That Reads Plain English Without a Transformer

Ada reads plain English through its own lexicon, learns from the decisions a business has already made, and shows its work, in milliseconds. Free in beta.

Glyphh AI today released Ada in beta, a classifier built on hyperdimensional computing instead of a large language model. It reads natural language without one. No transformer. No GPU. No training run.

Your history is the model. Record a decision once and Ada uses it on the very next one. Same question, same answer, every time.”

— Chris Timmerman, Founder, Glyphh AI

Most of the decisions a business makes aren’t new. Approve the loan or don’t. Route the message to this team or that one. Score the risk, flag the fraud, set the refund. Each call has been made hundreds of times before, and the outcome is on record. Today those decisions go to a person, or to a frontier model that reasons from scratch every time, pays for the privilege, and sometimes gives a different answer on a bad day.

Ada takes a third road. Describe your data once, record the decisions you’ve already made, and Ada classifies every new record from the ones it most resembles. Unlike a vector search, Ada compares your typed fields and your text together, learns from your own records which fields and words matter, and returns a calibrated probability and the cost of each choice, not a similarity score.

Marketing Technology News: MarTech Interview with Gil Bouhnick, Director of Product @ AppsFlyer

“Everybody’s building bigger brains to answer questions you’ve already answered a thousand times,” said Chris Timmerman, founder of Glyphh. “Your history is the model. Record a decision once and Ada uses it on the very next one. Same question, same answer, every time, and she shows you exactly which records she stood on. When she doesn’t know, she says so. I wish more AI did that.”

Natural language, no language model

Ada reads the natural language in your records, such as a ticket, a note or a claim description, with no language model in the loop. It runs in milliseconds, and your text never leaves for a model.

One record, many answers

A single record can answer many questions at once, each in its own type:

– Categories: which team, which product, which risk tier.
– Scores: ordered levels such as low to critical.
– Yes or no, with the probability of yes.
– Amounts, such as a refund or a price.
– Multiple labels, each with its own probability.
– Conditional questions, asked only when another answer calls for them.

Marketing Technology News: Segments Are Dead. Segmentation Has Never Mattered More.

How Ada works

– Learns from your own decisions. A new labelled record counts the moment it’s recorded. Nothing to retrain.
– Deterministic. The same input always gets the same answer.
– Shows its work. Every answer comes with calibrated probabilities and the precedents behind it.
– Knows its limits. When nothing on record is close enough, Ada hands the decision to a person or a language model, and that answer becomes a precedent for next time.
– Fast and cheap. About 6 milliseconds and a tiny fraction of a cent per decision.
– Private. Models run isolated per organization. Glyphh never trains on customer data.

Better in front of an LLM

Ada isn’t trying to replace language models. She sits in front of one, decides what she’s sure of, and passes the rest along. In Glyphh’s published bank-message routing test, Ada handled roughly four in five messages herself and handed the rest to Claude; the pair beat the model alone on accuracy at about a thousandth of the cost. Every result is public on GitHub.

“This is one person, a stack of high-dimensional vectors, and a lot of stubbornness,” said Timmerman. “A decision model shouldn’t get creative with your loan book. It should do what you’ve done before and show you the receipt. Build with AI. Run without it.”

Ada is named after Ada Lovelace, who wrote the first published program, for Charles Babbage’s Analytical Engine, about a century before the first electronic computer existed.