Teaching AI when a person could change the outcome of a message,
and how to bring them in.
Routine messages, an agent can send. The ones that can change a relationship need a person, and the hard part is knowing which, and where in the message. How you bring them in decides what you get: a quick answer, or their judgement.
The meeting notes say the CFO pushed back on price. A model reads that as a price objection. The writer, with the same notes, reads a CFO who was never going to say yes in front of his own team.
the context:
emails, meeting notes, history
the model
its read
their read
so they decide how hard to push, what to leave out, how much to say
We tested it on 9 Oct: records written by an open model didnโt lift its own scores on two social-reasoning tests. The read has to come from a person.
Labelled and expert data
Valued over $29bn after Metaโs investment, June 2025
Experts hired to make lab data
$10bn valuation, Oct 2025; $2bn+ gross run rate
Expertsโ step-by-step reasoning
Founded 2025; reported $3.2bn valuation, Sept 2026
AI-run expert network, London
$22.75m Series A led by a16z, May 2026
Social reasoning from real work, built to record outcomes
Pre-seed. Records start with the first cohort
A labโs own logs canโt capture it: an edit shows what changed, not why. What a person sees in the other person exists only as they write, inside the product they write with. That product is GLMMR.
How people read the person they write to.
How they read the person theyโre writing to, and what the relationship can take.
What they have seen happen before, never written down.
27 of 37
of the responses in our own early sessions drew on one of these: 15 on lived experience, 12 on the relationship.
Juliana Camargo
Co-founder and CEO
Ran her own consultancy, grown by word of mouth: 1,500+ one-to-one sessions drawing out the judgement of founders, CEOs and YPO leaders, and turning it into writing people trust. Moved to the UK from Brazil at 20, after placing top three nationally in a LโOrรฉal marketing competition.
Peter Jin Hong
Co-founder and Chief Product Architect
Fourteen years at Google and Alphabet, across Search, Assistant and Trust & Safety, working with behavioural science leaders Dr Dacher Keltner and the late Dr Albert Bandura. Speaks on trust and Human+AI collaboration, including at Harvardโs AI for Human Flourishing think tank.
Edwin Cubillos
Engineer (ex-Mercado Libre, Uber, Nokia)
Line Hjartarson
Design Lead / Lovable Ambassador (former Design Director: Omny, Cognite)
AI research scientist (name withheld)
Staff-level AI research scientist at a large technology company. Trains AI agents to read situations and follow the unwritten social rules of interacting with people. Interviewed and aligned; open to joining when the round closes.
Built for the messages at work where the relationship matters.
Every session is graded by hand against Julianaโs method: 200+ since July.
Where in the message human judgement is needed
How they were brought in
Their judgement: how they read the person theyโre writing to (social reasoning)
The reply they expected, and the one the AI expected
What came back
whose read
was right
Captured inside our product. Only writers who opt in; a lab gets labels, never the message or a name.
16 of 35
people in demos asked to use it on the spot.
340
on the waitlist, with no paid marketing.
after a demo
What it costs to make the data, at a planning price of $100 a record
Mercor, for example, pays its experts $60โ70 of every $100 it earns. Ours comes from a product people want to use for their work.
Mercor: Sacra; Dealroom, July 2026. GLMMR: our estimate of model spend, $3.30 to $4.20 a record, from first reads measured on test drafts ($0.19 to $0.41 each); a full session not yet measured.
500 records: the public test
the rest: for sale to labs
It measures how often the person beat the model on the same message. Any lab can run it. AfterQuery started the same way.
Launch. Records start.
Benchmark published. First lab pilot.
First lab contract, about $3M.
Three labs buying. About 90,000 records a year, from about 1,200 people who opt in.
Dates are targets.
It takes us to the published benchmark and the first lab pilot, spent in the order of the plan.
People
Owns the dataset, the benchmark and the statistics, and talks to labsโ data teams with Juliana. Then a platform engineer.
Data collection
Inviting and supporting the writers who opt in, and privacy counsel for consent and for any release of records to labs.
Model compute
Running the product for the cohort, a grader trained on hand grades, frontier models tested on our benchmark, and a tuned open model.
We build the product with AI coding agents and have merged 628 changes since July, so the round goes to research and data, not a large product team.
Juliana Camargo
Co-founder, CEO
Peter Jin Hong
Co-founder, Chief Product Architect