Cars in view — now
—
● From the library
—
—
The raw pixels
—
—
—
—
Reading the library…
Unclassified // Open source · derived from public NYC DOT frames · no faces · no full plates
Vehicle fingerprinting platform · Unclassified // Open Data
Count every street.
Read the highways.
We read New York's own traffic cameras in real time. Every camera gives an honest count — how many vehicles, what kind, how big. On the sharpest cameras, Bilbo reads further: the make, the color, the company on the side of the truck. And where the pixels can't back a claim, it says "can't tell" — never a guess. The traffic, never the people.
Reading the streets…
"Truth is a thing of this world… Each society has its régime of truth… the techniques and procedures accorded value in the acquisition of truth; the status of those who are charged with saying what counts as true."
— Michel Foucault, Truth and PowerThe instrument
917 cameras, on the ground
every public traffic camera in the five boroughs · colour is what its pixels can carry · click one to see what it watches
Click a camera to drop into it
What it can read
Blue — make, model, company
Orange — type, size, colour
Purple — counting only
Right now
The block, this second
live counts from the cameras · aggregate only
Cameras reporting live
—
fresh frames · stale feeds flagged, not counted
The honest ladder
917 cameras, three tiers
every camera was measured against real footage · a camera only claims what its pixels can carry
Blue — the readers
25
HD highway cameras · 1920px and up
- Everything below, plus in daylight: make, often the model, the company on the side, what it's carrying, racks and damage
- Which state the plate is from — by its color and design, roughly half the time
- Buses read almost like name-tags: route, destination, fleet number
- Never: the plate number, faces, or following a vehicle between cameras
Orange — the spotters
96
mid-res cameras · 640–1919px
- Everything below, plus daytime color and bus spotting
- Big-fleet recognition by color and shape — a hot-pink mover reads at 42 pixels
- Can't tell: make or model, fine lettering, plates
Purple — the counters
790
standard cameras · under 640px · all of Manhattan
- Count, type, size, direction — the honest floor, everywhere
- Daytime color, and loud standard liveries like taxi yellow
- Can't tell: make, model, company lettering, plates — the information isn't in the pixels
The constellation
917 eyes, drawing the city themselves
no map underneath — every dot is a real camera, placed by its own coordinates · the street grid appears on its own · hover any dot
blue · reads
orange · spots
purple · counts
reporting now
Time series
Traffic over time
avg cars in view · 1-min buckets
Rhythm
Rush curve — average by hour
all history · local time · when the block wakes and quiets
Composition
What's out there
all-time · vehicle mix by type
Measured, not guessed
Vehicle types by size
each camera learns its own scale from the typical car, then sizes every vehicle
The search layer
Handles — how a person asks the data
underneath, every vehicle gets a machine fingerprint no human can read · these are the plain-language handles built on top, so a person can ask "white Toyota with a ladder rack" or "Penske truck" or "the S78 bus" · each handle is only ever asserted where the camera tier can carry it
| Handle | Blue | Orange | Purple |
|---|---|---|---|
| Count · type · size · direction | ✓ | ✓ | ✓ |
| Color | day only | day only | day only |
| Vehicle class (sedan / SUV / box truck / semi…) | ✓ | ✓ | ✓ |
| Company / livery | day | big fleets, by color-block | loud liveries only |
| Make | day, usually | can't tell | can't tell |
| Model | day, clean angle | can't tell | can't tell |
| Plate state (never the number) | ~half, by design | can't tell | can't tell |
| Bus route · destination · fleet number | ✓ | route, often | bus only |
| Individual features (loads · racks · damage) | day | can't tell | can't tell |
"can't tell" is an answer, not a failure — a confident wrong label is the one thing this system refuses to produce.
The company log
Fleets on file
a living catalogue of the commercial fleets the cameras have read · once a livery is on file, the next sighting is a lookup, not a guess
| Company | Line of work | Livery | Vehicle | Confidence |
|---|---|---|---|---|
| opening the catalogue… | ||||
Personality
Fleet fingerprint — every corner's character
the vehicle type that runs highest here vs. the citywide baseline
| Intersection | Signature | Sedan | SUV/van | Pickup+ | Truck | Bus | Seen |
|---|
Leaderboard
Busiest intersections — all time
| Intersection | Vehicles | Passes |
|---|
Ground truth
By intersection — this pass
live per-camera vehicle counts · aggregate only
| Intersection | Tier | Cars | Trucks | Buses | Moto | mi |
|---|
BrainView · what the system noticed
The city, narrated.
Bilbo reads every corner and writes down what changed — surges, quiet streets, heavy-vehicle clusters, the busiest blocks on record. Plain-English dispatches, ranked by surprise, every one traceable back to the count.
Dispatches
What it's noticing
generated live from the last readings · every claim traces back to the data
Listening to the streets…
The Record
Go back on the record
every pass the system makes is committed to a public ledger, forever · pick a moment and read what it saw then — auditable down to the commit
opening the ledger…
By design
What we don't do
the ethical line, drawn inside the product itself
Bilbo does not follow a vehicle from one camera to the next, does not read plate numbers, and does not look at people at all. We built cross-camera matching once, measured it honestly — close to a coin-flip at these resolutions — and removed it rather than ship a guess dressed as a fact. Everything richer than a count (a make, a color, a company) is asserted only on cameras proven sharp enough to carry it, only in conditions where it holds, and below that bar the system says "can't tell." Findings, not footage.
Bilbo Academy · the eye, teaching itself · class in session 24/7
Watch it learn.
Every reading on this site comes from a vision model — and this is where those models go to school. Class runs continuously: five rotating GPUs, a fresh burst every 30 minutes on whichever key has the biggest bank of new footage, over and over until the curves flatten.
Estimated first plateau · New York time
—·—h
gathering pace — first samples land with the next harvest tick
Contacting the academy…
The loop
How the academy runs, continuously
the harvest never stops · the training never queues · every 30 minutes a fresh key runs
01 · HARVEST
Sweep the cameras
Round the clock, weighted by tier — the 25 blue cameras every sweep, oranges often, purples lightly. Rain, dusk and night get oversampled on purpose.
—
02 · LABEL
Name the easy, flag the hard
A stronger model auto-names the obvious. The sharpest, hardest close-ups go to the gold bench below for a careful read — or an abstention.
—
03 · TRAIN
Five keys, all firing
Every 30 minutes the key with the biggest bank of new crops runs one burst — all five heads, sequentially — then hands off to the next key in the rotation. No queue, no idle GPUs, no waiting for a nightly window.
—
04 · GRADE
Exam on unseen cameras
Every generation sits the same exam: cameras it never trained on, split day and night. A step up means real learning, not memorization.
—
05 · REPEAT
Until the curves flatten
Heads that pass earn the street. Heads that don't keep abstaining in public while they study in private. Bursts keep firing until the graph refuses to climb — that's the plateau above.
—
The corpus
What it studies · by camera tier
contacting the harvester…
The rotation
Five keys · one classroom
a burst fires every ~30 minutes on whichever key has banked the most fresh crops · below is the state of each key
The faculty
One head per skill — each on its own staircase
five heads, five keys, one classroom · every skill scored separately on cameras it never trained on · a head below its public bar abstains in public — "can't tell" is a passing answer
The company log
The catalogue, filling in
every fleet the gold judge reads off a truck's side becomes a permanent entry — the next sighting of that livery is a lookup, not a guess
By design
How this stays honest
the fine print, printed large
Every grade on this page is earned on cameras the student never trains on, split day and night, so a step up means real generalization — not memorization. Every skill is gated by camera tier: a head may only speak on footage proven sharp enough to carry its claim, and below that bar it must abstain — "can't tell" is a passing answer here, a confident wrong one is expulsion. Gold labels come from careful one-by-one reads of the actual pixels (the badge, the lettering, the livery), never from a model's hunch — the one auto-namer we auditioned called a clear Mercedes an "Acura", confidently, and was shown the door. And this page is watch-only: the rig publishes telemetry outward; nothing on the internet can reach back in.
About Bilbo Data
The ethical Palantir.
Truth about a city isn't neutral, and it isn't free-floating. It's produced — by whoever owns the instruments that measure the world. Own the instrument, own the truth. Bilbo Data takes one of those instruments out of the private room and hands it to everyone.
If you hold a gun and I hold a gun, we talk about the law.
If you hold a knife and I hold a knife, we talk about rules.
If we come empty-handed, we talk as equals.
But if you hold a gun and I hold nothing — there is no conversation.
— Michel FoucaultThe régime of truth
A handful of institutions are "charged with saying what counts as true" about the city. The state runs the cameras. Palantir and its cousins run the analysis. The facts get produced inside rooms you'll never enter, then handed down to you as objective. The instrument is the power.
"Truth is a thing of this world… Each society has its régime of truth… the techniques and procedures accorded value in the acquisition of truth; the status of those who are charged with saying what counts as true." — Michel Foucault, Truth and Power
Bilbo Data breaks that monopoly. The same class of instrument — a full Palantir-style eye on the streets, the same computer vision — pointed at the same city. The only difference is who holds it: everyone. Open data, open method, public the second it exists. Not a gentler tool, not a watered-down one — the real thing, taken out of the private room and put in the street.
The name — a tale of two Tolkien references
Both companies reached into the same book. They chose opposite ends of it.
Palantír · the Seeing Stone
What they named themselves after
The palantíri were the elves' crystal spheres that let you see across the world. By the Third Age they belonged to the powerful — Stewards, wizards, the Dark Lord — and they corrupted nearly everyone who looked in: Denethor to despair and suicide, Saruman to ruin. Even Sauron couldn't make the Stone lie; he could only show selective truths. An all-seeing instrument that rots the hand that holds it.
Bilbo · the Ring-bearer
What we named ourselves after
In the Ring-verse power is rationed to the elite — three rings for the Elven-kings, seven for the Dwarf-lords, nine for Mortal Men, one for the Dark Lord. Then a hobbit, the smallest and most overlooked creature alive, ends up holding the master ring itself — carries the ultimate instrument of power and stays himself, and becomes the first bearer ever to give it up freely.
That's the whole thesis: same seeing-stone, opposite hands. The instrument only kings and dark lords were meant to hold, put in the paws of the one creature that doesn't get ruled by it. And — literally — Bilbo is my girlfriend's dog: a small creature, no one's idea of a hero, who ends up carrying something the powerful would rather keep for themselves. Made anonymously, in New York.
From counting to reading
Bilbo began as a counter — how many cars cross a corner each minute — and on most of the city's cameras it still is one, proudly. But we measured all 917 cameras against real footage and found the fleet isn't uniform: twenty-five of them, all on the highways, are sharp enough to genuinely read a vehicle. On those, in daylight, Bilbo learns the make, often the model, the color, the company painted on the side, what's strapped to the roof — even which state the plate is from, by its color and design alone. A bus practically introduces itself: route, destination, fleet number.
Every one of those claims is tier-gated: a camera only asserts what its pixels can carry, and below that bar the answer is "can't tell" — printed on the page, as an answer, in public. That's the discipline that makes the reading trustworthy: we'd rather tell you nothing than tell you something almost true. What Bilbo still refuses to do is follow a vehicle from camera to camera, read a plate number, or look at a person. The eye that recognizes is the more powerful eye; the restraint is in what we refuse to let it see, even when we could.
CURB-OS
Rebuilding American parking supremacy
An operating system for the world's most contested eight feet of asphalt. Deployed at scale on President Street.
DOUBLE-PARK
Full-spectrum hazard awareness
Real-time detection of the delivery van that stopped exactly where it shouldn't. Mission-critical. Hydrant-adjacent.
SCHOOL-ZONE
Situational awareness for the 3 o'clock surge
When the bell rings, the corridor changes character. Bilbo sees the shift before the crossing guard does.
B-LINE
Autonomous corridor intelligence
Every bus, every block, timed camera-to-camera. Grandiose name. Humble mission. Counts the B44.
How we hold it without being corrupted — Privacy & Civil Liberties
- No faces. Never detected, never stored. At a traffic camera's resolution a person is a smudge, and we keep it that way.
- No plate numbers. At most, on the sharpest cameras, Bilbo reads which state a plate is from — by its color and design, the way any passenger would. The number itself is physically unreadable at these resolutions, and we never try.
- No tracking. Bilbo does not follow a vehicle from one camera to the next. Every reading belongs to a single camera and a single moment. There is no route, no history, no owner in our data — because we cannot, and will not, build one.
- Vehicles, not people. Bilbo counts and measures vehicles. It does not profile pedestrians — no shirt, no face, no name. People stay part of the crowd.
- "Can't tell" is an answer. Every claim is gated by what the camera can actually resolve — tier, distance, daylight. Below the bar, the system says so out loud instead of guessing. A confident wrong label is the one thing we refuse to ship.
- Findings, not footage. Every frame is read, then discarded. We keep the count and the shape of the traffic — not the video.
- Auditable. Public cameras in, open method, every claim stamped with a time and a confidence.