What is video intelligence?

Video intelligence is turning footage into structured data a system can act on: what happened, who or what was involved, and when. With bem you define the record you want as a JSON schema, send the video, and get that record back with a timestamp on each event and an audit trail on the run.

What kinds of video can bem read?

Any footage a person could watch and describe: CCTV and security cameras, thermal cameras, dashcams, drone and aerial footage, phone video, studio product video, broadcasts, and social video. The examples on this page cover ports, plants, patrols, perimeters, roads, fields, store aisles, returns desks, film scenes, sports, esports, news, and ads.

Which video formats and file sizes does bem accept?

MP4 and MOV. A workflow call accepts files up to 50 MB each, sent as a multipart upload or as base64 JSON.

Does bem use the audio as well as the picture?

Yes. Frames are sampled and the audio track is transcribed, so what is said and what is shown land in the same record. A price on a shelf tag and a quote to camera come back together.

Can I get timestamps for what happens in a video?

Yes. Ask for the moment in your schema and each event comes back with it. Every example on this page with an events list was produced that way, and a person can jump to the second and check.

How long does it take to process a video?

In the recorded runs on this page, a 12-second clip took about 5 to 9 seconds and two minutes of aisle footage took 22 seconds. Each example shows the time its run took.

Per function call, the same as any other file. There is no per-minute or per-frame meter, and the first 100 function calls are free.

How is this different from traditional video analytics?

Traditional video analytics ships a fixed detector for each task. With bem you describe the record you need in a schema, so the same footage can return a play index for one team and a sponsor report for another. The result then runs through the same workflow as your documents: Classify, Enrich, Decide, and Send, with an audit trail on every run.

Use cases · Video

Video intelligence for critical industries.

bem pulls the essential data out of the messiest information there is. Port cameras, patrol footage, broadcasts, and reels go in. Typed records come out, timestamped, with an audit trail on every run.

Start building free Talk to an engineer

Videos below: 18
Video sent to bem: 8 min 3 s
Time bem took: 2 min 34 s

7.6 slanded · The container lands securely onto the truck trailer bed.
4.8 scrouches_at_fence · thermal infrared camera
9.7 stransfer · rigid inflatable boat (rib) alongside fishing vessel
BX 4072plate.text · rear · hatchback, white
3.6 smedium · Front view of a vintage car parked in the pouring rain with headlights on.
9 spothole · left edge · high
00:33OT · 3rd & 7 · 16 - 23 · touchdown
$3.74Grade A Pasture Raised Large Brown Eggs · Marketside
8.99Circle Sign · shelf tag
00:11forecast · Lip-Bu Tan, Intel CEO
00:09A Kittle Goes A Long Way · ALUMINUM FREE
00:03blue stripe · light blue, brown

7.6 slanded · The container lands securely onto the truck trailer bed.
4.8 scrouches_at_fence · thermal infrared camera
9.7 stransfer · rigid inflatable boat (rib) alongside fishing vessel
BX 4072plate.text · rear · hatchback, white
3.6 smedium · Front view of a vintage car parked in the pouring rain with headlights on.
9 spothole · left edge · high
00:33OT · 3rd & 7 · 16 - 23 · touchdown
$3.74Grade A Pasture Raised Large Brown Eggs · Marketside
8.99Circle Sign · shelf tag
00:11forecast · Lip-Bu Tan, Intel CEO
00:09A Kittle Goes A Long Way · ALUMINUM FREE
00:03blue stripe · light blue, brown

01Critical operations

CCTV, dashcam, drone, and thermal video, read to the second.

The cameras critical industries already run. Press play, or let it run: each line of the record lights up as the video reaches it, and pressing a line takes the video there. The footage was generated for this page so it can be shown in full. The records are real bem runs on it.

01LogisticsTerminal operators, carriers

Generated footage

0:00 / 0:12

A crane move becomes a berth log.

A quay camera watches one container come off a ship and onto a truck. bem logs each phase of the move to the second, counts the people on the quay and what they are wearing, and says whether anyone was under the load.

The schema asks for

site
equipment[]
container{}
truck{}
events[]
moves_completed
personnel{}
people_under_load

What it feedsCrane productivity per berth, a time-stamped record of every move, and a safety check on each lift.

Quay camera, one container move. The footage was generated for this page; the record is what bem returned for it.

bem returns6.0 s

site: Container port terminal
equipment: Ship-to-shore gantry crane, Semi-trailer truck
container.length_class: 40-foot
moves_completed: 1
personnel.count: 2
personnel.ppe: Hard hat, High-visibility vest
personnel.position: Standing safely on the quay side away from the vehicle and container lift.
people_under_load: false

events[]
second	event	description
0	hoist	The crane hoists a brown container off a cargo ship.
2.5	traverse	The crane traverses horizontally carrying the container toward the waiting truck.
5	lower	The container is lowered down onto the flatbed trailer of the yellow truck.
7.6	landed	The container lands securely onto the truck trailer bed.
9.5	truck_departs	The truck drives forward and departs the loading zone.

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

02ManufacturingPlants, yards, vehicle logistics

Generated footage

0:00 / 0:12

A plant exit camera becomes a gate-out record.

An ordinary security camera over the exit lane: a finished car rolls past rows of wrapped stock and out the door. bem reads the plate off the moving car, describes the vehicle, notes the worker at the door and what he is wearing, and logs the moment it leaves the building.

The schema asks for

site
area
vehicle{}
plate{}
events[]
workers_visible
worker_ppe[]
door_open
exited_by_end

What it feedsA gate-out record for every unit by plate, matched to its build order, from cameras the plant already has.

CCTV, plant exit lane. The footage was generated for this page; the record is what bem returned for it.

bem returns4.8 s

site: factory storage warehouse
area: indoor vehicle holding area
plate.text: BX 4072
plate.position: rear
plate.plate_color: yellow background with black characters
vehicle.body_style: hatchback
vehicle.color: white
vehicle.visible_damage: false
workers_visible: 1
worker_ppe: high-vis vest
door_open: true
exited_by_end: true

events[]
second	event	description
8	passes_workers	The car passes the worker in the high-vis vest standing near the exit door.
11.2	exits_building	The vehicle drives through the open doorway and exits the building.

Recorded bem run, one Extract function. 2 of 3 events shown. Not every field is shown.

03DefenseNavies, coast guards, maritime security

Generated footage

0:00 / 0:12

Patrol footage becomes a contact log.

An aircraft holds its camera on a trawler sitting still on open water. A small boat runs in, comes alongside, and bales start changing hands. bem identifies both vessels, says which one is under way, and logs the approach, the moment they meet, and the transfer.

The schema asks for

view
sea_state
vessels[]
underway
events[]
alongside_at_end
transfer_observed

What it feedsA contact report written as the footage arrives, and a searchable log of every rendezvous across a season of patrols.

Aerial patrol camera, two vessels. The footage was generated for this page; the record is what bem returned for it.

bem returns6.6 s

view: high-angle aerial drone shot
sea_state: calm sea surface with gentle ripples
vessels[0].type: fishing vessel
vessels[0].description: blue hull with rusted areas, stationary with anchor line in water
vessels[0].underway: false
vessels[1].type: rigid inflatable boat (rib)
vessels[1].description: grey and white rib with outboard motor
vessels[1].underway: true
alongside_at_end: true
transfer_observed: true

events[]
second	event	description
0.8	second_vessel_appears	A small inflatable boat enters from the left side of the frame.
1.2	approaches	The inflatable boat approaches the stationary fishing vessel.
6.8	comes_alongside	The inflatable boat comes alongside the fishing vessel and a person prepares to board.
9.7	transfer	A person from the inflatable boat transfers items or boards the fishing vessel.

Recorded bem run, one Extract function. 4 of 5 events shown. Not every field is shown.

04DefenseBase security, critical infrastructure

Generated footage

0:00 / 0:12

A thermal camera becomes a perimeter event log.

Night, a fence line, an aircraft on the apron behind it. bem reads the thermal picture: two people come out of the trees, one goes to the fence and crouches at its base, the other hangs back and watches. Each step has its second.

The schema asks for

sensor
scene
persons_count
events[]
at_fence
behaviour
background_assets[]

What it feedsAn alert with the reason attached, and a log an operator can check in seconds, not a wall of monitors.

Thermal camera, airfield fence line. The footage was generated for this page; the record is what bem returned for it.

bem returns7.8 s

sensor: thermal infrared camera
scene: airport perimeter fence at night with airplane in the background
persons_count: 2
at_fence: true
behaviour: Two individuals approach an airport perimeter fence at night, with one crouching near the fence while the other stands watch in the background.
background_assets: perimeter chain-link fence with barbed wire, large airplane on runway, runway lights, trees and bushes

events[]
second	event	description
2.1	appears	Two figures appear near the bushes by the perimeter fence.
4	approaches_fence	One figure runs towards the fence line while the other stands back.
4.8	crouches_at_fence	The first figure crouches down at the base of the chain-link fence.
4.8	stands_watch	The second figure stands stationary in the background, keeping watch.

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

05InsuranceClaims, fleets, motor insurers

Generated footage

0:00 / 0:12

A dashcam clip becomes an incident account.

Rain, a residential street, a van coming out of a side road. bem gives the neutral account a claims handler needs first: the conditions, each vehicle and what it did, whether anything touched, and whether the car carrying the camera was involved. It does not assign fault.

The schema asks for

camera
road
weather
road_surface
vehicles[]
events[]
collision_occurred
camera_vehicle_involved
entered_from_side_road
injuries_visible

What it feedsA first notice of loss that fills itself in, and a fleet's near misses sorted without anyone watching the tape.

Dashcam, wet residential street. The footage was generated for this page; the record is what bem returned for it.

bem returns9.0 s

camera: Forward-facing dashcam mounted inside a vehicle traveling on a wet road.
road: Residential street with two lanes and a sidewalk on the right.
weather: Overcast and rainy
road_surface: Wet asphalt
vehicles[].label: orange hatchback, white van
entered_from_side_road: white van
collision_occurred: false
camera_vehicle_involved: false
injuries_visible: false

events[]
second	event	description
2.3	vehicle_ahead	The orange hatchback ahead applies its brake lights.
7.5	camera_vehicle_stops	The camera vehicle slows down and comes to a stop behind the stopped vehicles.

Recorded bem run, one Extract function. 2 of 3 events shown. Not every field is shown.

06InfrastructureRoad agencies, utilities, inspectors

Generated footage

0:00 / 0:12

A drone pass becomes a pavement survey.

A drone flies a stretch of two-lane road, looking straight down. bem lists each defect as it comes into view: what kind it is, where across the road it sits, and how bad it is. The pothole at the edge gets its own line.

The schema asks for

view
surface
defects[]
type
position
severity
defect_count
traffic_present

What it feedsA defect list per kilometre with the frame each one was found in, ready to become work orders.

Drone, looking straight down. The footage was generated for this page; the record is what bem returned for it.

bem returns6.6 s

view: Aerial view looking straight down at the road
surface: Asphalt with longitudinal cracking, transverse cracking, patches, and a pothole
defect_count: 5
traffic_present: false

defects[]
second	type	position	severity	description
0	longitudinal crack	centre line	high	Severe longitudinal cracking runs continuously along the center line of the road.
3	transverse crack	left lane	medium	A transverse crack branches outward from the center line towards the left edge.
8.5	patch	centre line and left lane	medium	A circular asphalt patch covers a repaired section across the center line.
9	pothole	left edge	high	A large pothole damages the pavement and edge near the left shoulder.

Recorded bem run, one Extract function. 4 of 5 defects shown. Not every field is shown.

07AgricultureGrowers, agronomists, crop insurers

Generated footage

0:00 / 0:12

A field flyover becomes a crop report.

A drone crosses a cornfield with one patch that is plainly not like the rest. bem names the crop, separates the healthy field from the stressed area, notes the irrigation rig, and says what someone should go and check on the ground.

The schema asks for

view
crop
planting
zones[]
condition
extent
overall_condition
equipment_visible[]
follow_up

What it feedsA scouting list for the agronomist, and evidence for a crop claim, from a flight that was already scheduled.

Drone survey, cornfield. The footage was generated for this page; the record is what bem returned for it.

bem returns6.6 s

crop: corn
planting: row crop
overall_condition: mostly healthy with a central anomaly
equipment_visible: pivot irrigation system
follow_up: Check soil moisture levels and irrigation coverage in the center section.

zones[]
second	condition	location	extent	description
0	healthy	surrounding field	large	The surrounding field consists of tall, healthy green corn plants growing in neat rows.
0	stressed	center cut-out section	medium	The central section has dried out or harvested corn stalks and exposed soil forming a shape.

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

02Retail, e-commerce, and media

Shelf audits, product listings, returns, and media catalogs, from video.

The same idea on the shop floor, in the product catalog, at the returns desk, and in a film library. Generated footage again, and real bem runs on it.

08Retail intelligenceStore operations, CPG field teams

Generated footage

0:00 / 0:12

An aisle walk becomes an out-of-stock list.

A phone moves along the juice aisle: full shelves, then a bay with almost nothing on it. bem names the category, finds each empty stretch, says which shelf it is on and what sits next to it, and notes that the price labels are still there.

The schema asks for

store_area
category
gaps[]
shelf
width
neighbours
gap_count
price_labels_present
facing_quality

What it feedsA restock task while the walk is still going, and on-shelf availability by store without a manual count.

Phone walk, beverage aisle. The footage was generated for this page; the record is what bem returned for it.

bem returns6.2 s

store_area: Beverage aisle
category: Juice
gap_count: 4
price_labels_present: true
facing_quality: Neatly aligned on stocked sections

gaps[]
second	shelf	width	neighbours	description
1.5	second from top	wide	berry juice bottles	A wide empty section on the second shelf from the top.
2.1	third from top	wide	berry juice bottles	A large empty gap spanning most of the third shelf from the top.
3.2	bottom	wide	orange and blue juice bottles	An extensive empty area on the bottom shelf.

Recorded bem run, one Extract function. 3 of 4 gaps shown. Not every field is shown.

09E-commerceMarketplaces, catalog teams

Generated footage

0:00 / 0:12

A product video becomes a listing.

A seller uploads a turntable video and nothing else. bem returns the product type and where it belongs in the catalog, the colours, materials and closure it can see, which angles the video covers, and a draft title and bullets with no brand invented.

The schema asks for

product_type
category_path
colors[]
upper_material
sole
closure
branding_visible
views[]
suggested_title
suggested_bullets[]

What it feedsA catalog entry drafted from the video alone, ready for a person to approve.

Studio turntable, running shoes. The footage was generated for this page; the record is what bem returned for it.

bem returns5.3 s

product_type: Sneakers
category_path: Shoes > Athletic > Running
colors: white, orange, grey
upper_material: mesh
sole: thick rubber foam sole
closure: lace-up
branding_visible: false
suggested_title: White and Orange Athletic Running Shoes
suggested_bullets[0]: Breathable mesh upper with vibrant orange side accents
suggested_bullets[1]: Max-cushion platform sole for impact absorption
suggested_bullets[2]: Secure lace-up closure with padded tongue

views[]
second	view	description
0	side, toe pointing right	Side profile of the right shoe showing its thick white sole and orange accent stripe.
1.2	three-quarter	Three-quarter front view of the pair of shoes on the rotating platform.
3.7	front (toe)	Front view focusing on the toes and laces of both sneakers.

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

10E-commerceReturns, fulfilment, marketplaces

Generated footage

0:00 / 0:12

A returns desk becomes an inspection record.

An overhead camera watches a parcel come back. bem follows the hands: the box opens, a kettle comes out, a finger traces the crack down its body. The record says what the item is, what came with it, where the damage is, and that it cannot go back on the shelf.

The schema asks for

station
item{}
parts_present[]
packaging
condition{}
events[]
resellable

What it feedsA refund decision with the evidence attached, and a grade for every returned unit.

Overhead camera, returns desk. The footage was generated for this page; the record is what bem returned for it.

bem returns5.9 s

station: Inspection workstation
item.type: Kettle
item.color: White
parts_present: Kettle, Power base
packaging: Cardboard box with inner packaging
condition.damaged: true
condition.damage_type: Crack
condition.damage_location: Front body of the kettle
condition.severity: functional
resellable: false

events[]
second	event	description
0	opens_parcel	The worker opens the cardboard box parcel.
2.6	removes_item	The worker removes the kettle from the box.
4.1	inspects_item	The worker holds up the kettle to inspect its exterior.
6.1	indicates_damage	The worker points out the prominent crack on the kettle's body.
10.4	sets_down	The worker sets the kettle down on its power base.

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

11Media & entertainmentStudios, streamers, archives

Generated footage

0:00 / 0:12

A film scene becomes catalog labels.

Twelve seconds of a noir: a man in a trench coat on a wet street, a car with its lamps on, a woman in a hat. bem returns what a catalog needs to file it: genre and sub-genres, era, mood, a shot list with the second each cut lands, who is in it, advisories, keywords, and a logline. It also catches that the picture is not purely black and white.

The schema asks for

content_type
genre
sub_genres[]
era_depicted
color
mood[]
shots[]
characters[]
dialogue_present
content_advisories{}
keywords[]
logline

What it feedsCatalog metadata, search, and content classification across a library nobody has time to watch.

Film scene, three shots. The footage was generated for this page; the record is what bem returned for it.

bem returns6.8 s

content_type: feature film scene
genre: Mystery
sub_genres: Noir, Thriller, Crime
era_depicted: 1940s
color: mixed
mood: suspenseful, moody, mysterious, gritty
setting: Rain-slicked city streets at night
shot_count: 3
characters[0]: A person in a fedora and trench coat walking
characters[1]: A woman wearing a wide-brimmed hat with dark clothing and red lipstick sitting inside a car
dialogue_present: false
content_advisories.violence: none
content_advisories.weapons_visible: false
keywords: noir, rain, vintage car, fedora, mysterious, detective, black and white, night
logline: A mysterious figure walks a wet, dark city street before a woman waits in a vintage car during a rainstorm.

shots[]
second	shot_type	description
0	wide	A figure in a trench coat and fedora walks down a rainy, dark city street.
3.6	medium	Front view of a vintage car parked in the pouring rain with headlights on.
6.5	close-up	A woman in a wide-brimmed hat and dark clothing sits inside a car.

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

03Seven real posts

Sports, esports, news, and social video: real posts, real records.

Nothing here was staged. Each one is a public post, sent to bem as it is: cuts, captions, crowd noise, and shaky hands included. Watch it, then read what came back.

12EsportsTournament organisers, teams, data providers

A tournament round becomes a scoreboard record.

Nineteen seconds of a Counter-Strike broadcast: one player's view, a scoreboard across the top, a row of player cards along the bottom, then a cut to his face and the crowd. bem reads the round and the score off the overlay, sees that the bomb is down, names the player being watched and his team, and marks where the broadcast cuts away.

The schema asks for

game
map_site
hud{}
observed_player
observed_player_health
team_left_players[]
segments[]
sponsors_on_overlay[]

What it feedsRound-by-round match data, a searchable highlights archive, and sponsor exposure, straight from the broadcast.

BLAST Premier on YouTubeCounter-Strike 2 tournament broadcast, 0:19. The first 19 seconds of a longer highlights video.

bem returns6.2 s

game: Counter-Strike 2
hud.round: ROUND 18 / 24
hud.score_left: 11
hud.score_right: 6
hud.bomb_planted: true
map_site: A
observed_player: NiKo
observed_player_health: 91
team_left_players: TeSeS, kyousuke, karrigan, m0NESY, NiKo
sponsors_on_overlay: Corpay, GamingMalta

segments[]
second	kind	description
0	gameplay	NiKo holding an angle with an AK-47 on site A
16.2	player camera	NiKo in the booth wearing headset
17.7	crowd	Crowd cheering and clapping in red lighting

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

13Sports dataLeagues, broadcasters, betting

@nflTop 5 plays of Week 9, 0:56.

A top-five countdown becomes a play-by-play table.

The league's own Week 9 countdown: five plays cut from five broadcasts, each with a different score bug. bem reads the rank, the score, the quarter, the clock, and the down off the screen for every play, and says what happened.

The schema asks for

league
presented_by
plays[]
rank
score_on_bug
quarter
game_clock
down_and_distance
play_type
touchdown
description

bem returns24.0 s

league: NFL
play_count: 5
presented_by: Jersey Mike's Subs
plays[0].matchup: TB at CHI

plays[]
timestamp	rank	score_on_bug	quarter	game_clock	down_and_distance	touchdown	description
00:00	5	TB 0 - CHI 7	1st	6:10	3rd & 8	true	A player rushes through defenders and breaks free into the end zone for a touchdown.
00:10	4	0 - 0	1st	5:43	3rd & 23	true	A deep pass is completed and the receiver breaks away down the sideline for a touchdown.
00:24	3	13 - 14	3rd	1:32	1st & GOAL	false	An attempted reception in the end zone results in a fumble near the goal line.
00:33	2	16 - 23	OT	1:08	3rd & 7	true	A receiver catches a pass over the middle and dashes down the field for a long touchdown.
00:43	1	24 - 9	3rd	8:20	3rd & 6	true	A receiver makes a contested catch in the end zone right at the pylon for a touchdown.

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

What it feedsA searchable play index, a highlights feed, or a sponsor report, built from footage that already aired.

14Retail intelligenceBrands, category and insights teams

@bobbyparrishCreator product roundup, Walmart, 1:39.

A creator's grocery run becomes a price and claims sheet.

A creator walks a Walmart and recommends five products to camera. bem returns each one with its brand, the price on the shelf tag, the claims printed on the pack, and the reason he gave, in his words.

The schema asks for

retailer
format
products[]
price
pack_claims[]
why_recommended
quote
other_brands_on_shelf[]
call_to_action

bem returns7.5 s

retailer: Walmart
format: creator product roundup
products[0].quote: “This dozen is cheaper than the organic non-pasture raised eggs. Epic deal.”
other_brands_on_shelf: Bettergoods, Dave's Killer Bread, Purely Elizabeth, Bare, Cheerios, Kashi
call_to_action: “Make sure to give me a follow so you never miss out on really good, good bobby approved food.”

products[]
price	product	brand	pack_claims	why_recommended
$3.74	Grade A Pasture Raised Large Brown Eggs	Marketside	Pasture Raised · CA Cage Free · Farm Fresh	It's pasture-raised, certified humane, and cheaper than the organic non-pasture-raised eggs.
$7.26	Grass-Fed Ground Beef 80% Lean / 20% Fat	Marketside	100% Grass-Fed & Finished · Beef Raised with No Antibiotics or Added Hormones EVER	It's 100% grass-fed beef with 80/20 fat ratio, which provides more flavor and is cheaper than the leaner 93/7 option.
$13.82	California Extra Virgin Olive Oil	Great Value	100% California · First Cold Pressed · Smooth & Balanced	It is single origin from California, comes in a glass bottle, and has an unbeatable price per ounce.
$5.92	Stone Milled Mighty Wheat Organic Bread	One Mighty Mill	Organic Bread · No Seed Oils · 5g Plant Protein · No Refined Sugar	It uses organic grains, EVOO, and stone-milled flour to make nutrient-dense bread.

Recorded bem run, one Extract function. 4 of 5 products shown. Not every field is shown.

What it feedsA feed for brand and category teams: which products creators recommend, at what price, and why.

15Shelf auditsRetail operations, merchandising

@aldi.mademedoitAisle walk, Aldi seasonal finds, 9:09. bem was sent the first two minutes.

A phone walk down the aisle becomes a shelf record.

Someone films the seasonal aisle at Aldi, shelf by shelf. bem lists each product the camera stops on, reads the price off the shelf tag, and says where each price came from.

The schema asks for

retailer
section
seasonal_theme
items[]
brand
shelf_price
price_source
items_with_price

bem returns22.1 s

retailer: Aldi
section: Home Finds / Christmas
seasonal_theme: Christmas

items[]
timestamp	product	brand	shelf_price	price_source
00:12	5 FT. (152 cm) LED Flocked Alpine Tree	Merry Moments	null	none
00:23	Holiday Food Storage	Merry Moments	5.99	shelf tag
00:34	Circle Sign	Merry Moments	8.99	shelf tag
00:51	Kitchen Mat	Merry Moments	6.99	shelf tag
01:26	Gel Clings	Merry Moments	1.99	shelf tag
01:30	Mirror House Candle	Merry Moments	5.99	shelf tag

Recorded bem run, one Extract function. 6 of 12 items shown. Not every field is shown.

What it feedsPrice checks, assortment tracking, and planogram audits from footage a phone already took.

16News monitoringResearch desks, comms, compliance

@yahoofinanceInterview clip, Intel's CEO, 0:34.

A CEO interview clip becomes a list of statements.

A 33-second cut from a finance interview. bem names the speaker from the lower-third graphic, reads the headline on screen, and returns each claim with a verbatim quote and the moment it was said.

The schema asks for

outlet
headline_on_screen
speakers[]
topic
statements[]
stance
forward_looking
companies_mentioned[]

bem returns12.3 s

outlet: Yahoo Finance Invest
headline_on_screen: INTEL CEO: AI INVESTORS NEED TO BE MORE PATIENT
speakers[0].name: Lip-Bu Tan
speakers[0].title: Intel CEO
speakers[0].identified_by: lower-third graphic
topic: AI investments and patient capital
forward_looking: true

statements[]
timestamp	type	claim	quote
00:00	fact	AI capital requirements are extremely large in the scale of trillions of dollars.	“The capital requirement is humongous. And then the trillion of dollars.”
00:11	forecast	AI investors need to be patient because returns will take longer to materialize.	“I think the money spent is well spent, but the return will be slower to come in. So people have to be more patient…”
00:27	opinion	Debt financing structures for AI are becoming challenging.	“So in term of the debt financing structure, financing, those become challenging.”

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

What it feedsMedia monitoring and research: who said what, when, and in their own words.

17Ad analysisMarketing, legal, competitive intel

@oldspiceProduct ad, NFL collection, 0:16.

A 16-second ad becomes a claims inventory.

A product spot with no dialogue to lean on: four packs and an end card. bem returns each product with its format, the scent printed on it, the player it is themed around, and every claim on the pack.

The schema asks for

advertiser
campaign_line
licensed_partner
products[]
on_pack_claims[]
claims[]
end_card
call_to_action
disclaimers[]

bem returns5.6 s

advertiser: Old Spice
campaign_line: THE GREATEST SMELL IN THE NFL
licensed_partner: NFL
claims: 24/7 PROTECTION · SHOWER 24/7 FRESH · ALUMINUM FREE
disclaimers: [ ]
call_to_action: null
duration_seconds: 15.75

products[]
timestamp	name	format	scent	featured_person	on_pack_claims
00:01	Shake & Break	dry spray	Swagger + Spin Moves	Bijan Robinson	24/7 PROTECTION · ANTIPERSPIRANT
00:03	Catch Me If You Lamb	body wash	Deep Sea + Dimes	CeeDee Lamb	SHOWER 24/7 FRESH
00:06	Saquon Is On One	dry spray	Santal + Speed	Saquon Barkley	24/7 PROTECTION · ANTIPERSPIRANT
00:09	A Kittle Goes A Long Way	deodorant stick	Musk + Moving Chains	George Kittle	ALUMINUM FREE

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

What it feedsA creative library, competitive tracking, and claims review before an ad runs.

18E-commerceMarketplaces, catalog teams

@targetdoesitagainTry-on, four colourways, 0:07.

A seven-second try-on becomes a product listing.

A shopper films four colourways of one fleece in a store mirror. bem returns the garment, each variant with its colours and how it was shown, and a draft title and bullets grounded in what is visible.

The schema asks for

garment_type
department
retailer
variants[]
fabric
fit
styled_with[]
suggested_title
suggested_bullets[]
price_shown

bem returns4.5 s

garment_type: quarter-zip fleece pullover
department: women's apparel
retailer: Target
fabric: fuzzy sherpa fleece
styled_with: medium wash straight-leg jeans, dark loafers
suggested_title: Women's Patterned Sherpa Fleece Quarter-Zip Pullover
suggested_bullets[0]: Soft sherpa fleece fabric keeps you cozy and warm
suggested_bullets[1]: Convenient quarter-zip front closure with contrast trim
price_shown: null

variants[]
timestamp	pattern	colors	shown
00:00	zebra stripe	black, beige, red	worn
00:01	pink checkerboard	pink, white	on hanger
00:03	blue stripe	light blue, brown	both
00:05	red floral	off-white, red	worn

Recorded bem run, one Extract function. All rows shown. Not every field is shown.

What it feedsA product catalog: a try-on video becomes listing attributes and draft copy for review.

These are public posts, shown with Instagram's and YouTube's own embeds and credited to their publishers. bem is not affiliated with or endorsed by any of them.

04One workflow

From video to a decision, in one workflow.

Extract turns the video into a record. The rest of the platform takes it from there: match it to your data, apply your policies, and send the result where the work happens.

Input

VIDVideoMP4 or MOV
AUDIts audio trackTranscribed with the picture

Components

01ClassifyKnows an ad from an interview from an aisle walk, and sends each the right way.
02ExtractSamples the frames, transcribes the audio, and returns the fields in your schema.
03EnrichMatches what was seen to your catalog, roster, or price file.
04DecideApplies your policies to the record.
05SendDelivers the result to your system.

Decision

Send to claims review

The ad makes three product claims (24/7 PROTECTION, SHOWER 24/7 FRESH, ALUMINUM FREE) and shows no on-screen disclaimer. Policy: a claim with no disclaimer goes to review.

Claims and disclaimers are from the recorded run on the Old Spice ad above. The policy and the decision are an example.

How the platform works Decide: your policies, applied

05For developers

The video extraction API: a schema and a file.

Define the record you want once. Then send video to the same endpoint you would send a PDF to. The response is your schema, filled in.

terminalbem API v3

# 1. Say what you want back
curl -X POST https://api.bem.ai/v3/functions \
  -H "x-api-key: $BEM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "functionName": "ad-breakdown",
    "type": "extract",
    "outputSchemaName": "AdBreakdown",
    "outputSchema": {
      "type": "object",
      "properties": {
        "advertiser":    { "type": "string" },
        "campaign_line": { "type": "string" },
        "products": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "timestamp": { "type": "string", "description": "MM:SS" },
              "name":      { "type": "string" },
              "format":    { "type": "string" },
              "on_pack_claims": { "type": "array", "items": { "type": "string" } }
            }
          }
        },
        "disclaimers": { "type": "array", "items": { "type": "string" } }
      }
    }
  }'

# 2. Send the video
curl -X POST https://api.bem.ai/v3/workflows/ad-breakdown/call \
  -H "x-api-key: $BEM_API_KEY" \
  -F "wait=true" \
  -F "file=@spot.mp4"

responsetransformedContent

{
  "advertiser": "Old Spice",
  "campaign_line": "THE GREATEST SMELL IN THE NFL",
  "products": [
    {
      "timestamp": "00:01",
      "name": "Shake & Break",
      "format": "dry spray",
      "on_pack_claims": ["24/7 PROTECTION", "ANTIPERSPIRANT"]
    },
    {
      "timestamp": "00:03",
      "name": "Catch Me If You Lamb",
      "format": "body wash",
      "on_pack_claims": ["SHOWER 24/7 FRESH"]
    },
    {
      "timestamp": "00:06",
      "name": "Saquon Is On One",
      "format": "dry spray",
      "on_pack_claims": ["24/7 PROTECTION", "ANTIPERSPIRANT"]
    },
    {
      "timestamp": "00:09",
      "name": "A Kittle Goes A Long Way",
      "format": "deodorant stick",
      "on_pack_claims": ["ALUMINUM FREE"]
    }
  ],
  "disclaimers": []
}

Read the quickstart

06More to build

Video data you can rely on.

01
Your schema, not ours
You decide the fields. The same video returns a play index for one team and a sponsor report for another.
02
Picture and sound together
Frames are sampled and the audio track is transcribed, so a price on a shelf tag and a quote to camera land in one record.
03
A timestamp on what happened
Ask for the moment and each event comes back with it, so a person can jump to the second and check.
04
An audit trail on every run
What went in, what came out, and which version of the function produced it. Open any result and trace it to its input.
05
The same workflow as documents
Video is one more input. Classify, Enrich, Decide, and Send work on it exactly as they do on a PDF.
06
Priced per call
A video costs the same per function call as any other file. There is no per-minute meter.

08Questions

Questions about video intelligence.

What is video intelligence?: Video intelligence is turning footage into structured data a system can act on: what happened, who or what was involved, and when. With bem you define the record you want as a JSON schema, send the video, and get that record back with a timestamp on each event and an audit trail on the run.
What kinds of video can bem read?: Any footage a person could watch and describe: CCTV and security cameras, thermal cameras, dashcams, drone and aerial footage, phone video, studio product video, broadcasts, and social video. The examples on this page cover ports, plants, patrols, perimeters, roads, fields, store aisles, returns desks, film scenes, sports, esports, news, and ads.
Which video formats and file sizes does bem accept?: MP4 and MOV. A workflow call accepts files up to 50 MB each, sent as a multipart upload or as base64 JSON.
Does bem use the audio as well as the picture?: Yes. Frames are sampled and the audio track is transcribed, so what is said and what is shown land in the same record. A price on a shelf tag and a quote to camera come back together.
Can I get timestamps for what happens in a video?: Yes. Ask for the moment in your schema and each event comes back with it. Every example on this page with an events list was produced that way, and a person can jump to the second and check.
How long does it take to process a video?: In the recorded runs on this page, a 12-second clip took about 5 to 9 seconds and two minutes of aisle footage took 22 seconds. Each example shows the time its run took.
How is video priced?: Per function call, the same as any other file. There is no per-minute or per-frame meter, and the first 100 function calls are free.
How is this different from traditional video analytics?: Traditional video analytics ships a fixed detector for each task. With bem you describe the record you need in a schema, so the same footage can return a play index for one team and a sponsor report for another. The result then runs through the same workflow as your documents: Classify, Enrich, Decide, and Send, with an audit trail on every run.

See pricing Supported file types

Send your first video.

The first 100 function calls are free. Pricing is per task, not per token or page. Pay as you go, with no lock-in.

Start building free Talk to an engineer See pricing

Video intelligence for critical industries.

Input

Components

Decision

Full-motion video review

Maintenance and inspection video

Body-worn and dash camera footage

Infrastructure inspection

Claim walkarounds

Safety camera review

Dock and yard footage

Earnings and investor video

Recorded visits and training

Seller and listing video

Sponsorship and brand exposure

Sponsored content review

Your schema, not ours

Picture and sound together

A timestamp on what happened

An audit trail on every run

The same workflow as documents

Priced per call

What is video intelligence?

What kinds of video can bem read?

Which video formats and file sizes does bem accept?

Does bem use the audio as well as the picture?

Can I get timestamps for what happens in a video?

How long does it take to process a video?

How is video priced?

How is this different from traditional video analytics?

Send your first video.