Phase 1 Β· Meet Tabi
$200/night
drag to procrastinate βweek 0.0
Shibuya Stream Hotel
$200/ night
Drag the weeks forward β prices climb, the chat goes quiet, and Tabi rots. Book to revive her.
Everyone wants to go to Japan.
Nobody books it.
Meet Tabi β the group-chat pet that dies if your trip doesn't happen.
Add to Telegram@PetSamaBot
How it works
Tabi joins your group chat, listens to everything, runs Tinder-style votes on everyone's picks, and iterates until the whole group agrees.
The brain
A team of Gemini agents on LangGraph β plus a model we trained ourselves.
The blocker agent is FreeSolo β QwenΒ 4B, fine-tuned with GRPO. On a held-out eval it beats a frontier model at finding the group's real blocker and speaking in the pet's voice:
Pulls constraints from text, voice notes, and images.
Melts everyone's price, location, and time into one plan.
Finds who's stalling the trip β and calls them out by name.
A team of agents, run as one deterministic supervisor.
We trained the pet's brain ourselves
Our 4B model vs. a frontier model
held-out eval Β· 50 examples Β· higher is better
Our 4B (live) Frontier (zero-shot)
Gold-F1closeness to the ideal reply
Our 4B
0.317
Frontier
0.078
In-voicesounds like the pet, not an assistant
Our 4B
98%
Frontier
0%
Trained on data the product generated about itself β SFT β distillation β GRPO on QwenΒ 4B β and the live pet runs on it right now.
Eight agents on the graph: supervisor Β· stage tracker Β· profile tracker Β· messenger Β· brain Β· phoebe Β· greenplanner Β· face
See it live β
This is where we open Telegram and plan a real trip. Watch the video:
Green by default
Every booking is carbon-scored the moment it's on the table.
Flights, trains, hotels, even the airport transfer β priced in COβe from published government factors, computed locally, so the numbers are impossible to inflate.
Flights
DEFRA β incl. radiative forcing
Trains
JR Central β Shinkansen factors
Hotels
by class & rooms the party needs
Transfers
airport β hotel last mile
the group's shared ledger
142miles
not driven
not driven
greener flight picked
low-carbon stay won the vote
rail routed over cars
every green choice, counted and totalled
We never log a footprint. We log a difference.
βYour trip emits 3,080 kgβ is true and useless β it isn't a saving. Every entry answers a narrower question: versus the option actually on the table, how much less is this?
Nonstop YYZβNRT, 10,300 km3,079.6 kg
Via Chicago +475 km, +1 landing3,321.7 kg
Avoided per person242.1 kg
Γ 4 travellers968.4 kg
Sources: DEFRA Β· EPA Β· JR Central. Baseline is the median of your own shortlist β savings only count when your pick beats it. Every figure computed from the live module.
Show your work Β· the carbon math
The ledger never records a footprint. It records a difference.
If a choice isn't better than the alternative on the table, nothing is recorded β carbon accounting that would survive an auditor.
Distance or nights Γ a published factor β every constant local, no API in the loop.
The dirtiest flight listed, the deck-average hotel, the same trip by car.
Four people flying is four times the flight β factors scale by group size.
A worse pick earns zero, never a negative. Bad choices can't grow the ledger.
Not a slide β that's the live product. Type /saved in the chat and Tabi computes the group's real ledger on the spot, from the same module behind every figure on this page.
Factors: DEFRA 2024 Β· CHSB 2023 Β· EPA Β· JR Central
Your souvenir
Book it, and Tabi mints your trip's Spotify Wrapped.
A permanent memory card for the group β the destination, how many iterations and how long it took to book, and the carbon you avoided. Minted once on Solana, so it's immutable. Yours.
Trip Wrapped2026
Kyoto Β· Aug 1β5
π DestinationKyoto
π Iterations to book12
β±οΈ Time to book3 days
π± Carbon avoided142 mi not driven
minted on Solana devnet Β· immutable Β· one per groupverify
Stop letting the trip die in the group chat.
Give your friends something that gets sad when you don't book β and a spring morning in Kyoto when you do.
Add to Telegram@PetSamaBot