INSIDE THE SYSTEMLLSTM / LARGE LOGISTICS SPATIO·TEMPORAL MODEL
Not one model. A system that learns the city.
LLSTM is a collection of systems that see the city as it is and as it is about
to be, act on it, and grade themselves before anything reaches a real rider.
Every decision is written down — so whatever intelligence we plug in next
starts smarter than the last.
VIEWS what it sees
↓ AGENTS what decides
↓ LEVERS how it acts
↓ THE CITY responds
↓ THE LOG remembers ↶ and the views learn
One loop, running continuously: agents read the views, pull the
levers, the city responds, the log records it — and the views learn. The rest of
this page walks it, station by station.
00 · THE PROBLEMOPEN LOOP / DEMAND WE DON'T CONTROL
The city changes overnight.
Demand reaches our network from brands we don't control. We don't set their
prices, run their promotions, or choose where they open their next store — and
any one of those can redraw the city's demand in a day. A model trained on last
month's city is stale the moment a brand changes its mind.
Platforms that own their app, their prices, and their promotions never had this
problem, so their playbooks don't transfer. An original problem needs an original
system. LLSTM is built for exactly this: a city that keeps changing under it.
01 · VIEWSWHAT THE SYSTEM SEES
The city, as it is and as it's about to be.
Everything the system knows is held in views — legible, checkable pictures of
the city, refreshed continuously. A forecast you can read is a forecast you can
grade, and every view is graded against what the city actually did.
Present
Every rider, every order, every store — live. When operations report a
store down, one update flows into every decision that follows.
Future
Demand forecast at several horizons and several scales — the city's next
hour, before it happens. The wide view gives the close view its context.
Past
What this corner, this hour, this weekday usually looks like — the
baseline the present is read against. Friday evening is not Tuesday evening.
02 · LEVERSHOW THE SYSTEM ACTS
Two ways to act on a city.
The system holds levers of two kinds: actions it can take in the world, and
knobs — the tunable weights inside its own fast decisions.
Actions
Guide a rider toward where the next order is likely to be. Open slots where
the work will land. Reach riders where the fleet is running thin.
Knobs
Every fast decision balances competing pulls — the fare in hand against
where the drop leaves the rider for the next one. That balance is a knob, and
it moves as the city does.
A longer trip pays more now; a shorter one ends next to the evening's busiest
corner. Neither answer is always right — the weighting shifts with the weather,
the hour, the fleet. Tuning it is a decision too, and it is logged like one.
03 · AGENTSWHAT SITS BETWEEN VIEWS AND LEVERS
Progressively smarter tenants.
Between the views and the levers sits whatever intelligence we trust most
today. The agent is replaceable by design — the views, the levers, and the log
are permanent, and every successor inherits them.
RUNNING
Scored models
Millisecond decisions, thousands a day: which order, which rider, where to
wait. Fast, consistent, and graded on every one.
NEXT
A supervisor
Reads every view, turns the knobs, responds when the world shifts — a
store down, weather turning, a quiet evening. Slow decisions, made well.
THE HORIZON
A native model
Trained on the network's own record until the city reads like language —
and deciding becomes prediction.
04 · EVALUATIONPROOF BEFORE THE STREET
Nothing ships on sounding right.
The system grades itself at three levels, and a policy that can't pass doesn't
reach a rider.
Grade the views
Was the forecast right? Every prediction is scored against what the city
actually did.
Grade the actions
Did the world move the way the decision assumed? New policies replay
against a digital twin of the city before they touch the street.
Grade the whole
Is a rider's hour worth more? The only score that finally matters.
And when the system is wrong, riders don't pay for it. That promise is built
into the product — not written in a policy page.
05 · THE LOGTHE ASSET THAT COMPOUNDS
Models are replaceable. The record is not.
Every decision writes a row: what the city looked like, what we did, what
happened. The log is three assets at once.
Sharper views
Every outcome is a label. The forecasts retrain on what actually happened,
and every view gets sharper.
Better decisions
The record keeps deliberate experiments alongside the everyday —
evidence about the roads not taken, which is what learning better decisions
requires.
Safe replacement
A candidate policy is replayed against the record before it ships. Most of
the risk of change becomes an offline computation.
An architecture can be copied the day it's published. A record of a city's
decisions and their outcomes can only be earned the way we earn it: by operating,
one day at a time.
THE DESTINATION
A model that speaks the language of mobility.
Views, levers, outcomes — one vocabulary. One model of the city, learning
continuously, deciding in milliseconds.
WHO ARE YOUINVESTORS / ENGINEERS
Investors
The models are replaceable. The loop and the record are not — and they
compound every week the network runs.