An example

You get clean data, shaped to your schema.

You tell Lirovo the fields you want, and it returns typed JSON that matches them. Every value carries a confidence score, and every value links back to the exact second and modality it came from. Here is what that looks like on a real video.

Define a schema, get a typed record back

This run is a Mistral AI CEO interview on CNBC. On the left, the fields we asked for. On the right, the typed JSON Lirovo returned, conforming to exactly that shape.

schema.json

{
  "speakers": "string[]",
  "organizations": "string[]",
  "funding_round": {
    "investor": "string",
    "stake": "string",
    "amount": "string"
  },
  "valuation": "string",
  "topics": "string[]"
}

output.json

{
  "speakers": ["Arthur Mensch"],
  "organizations": ["Mistral AI", "ASML", "Nvidia"],
  "funding_round": {
    "investor": "ASML",
    "stake": "~11%",
    "amount": "$2B"
  },
  "valuation": "$14B",
  "topics": ["semiconductors", "AI compute"]
}

The schema is enforced. Lirovo validates the output against the JSON Schema you give it, so what comes back conforms to what you asked for. In this run the filled values land between 0.92 and 0.97 confidence, and each one keeps an evidence pointer to the moment it was drawn from.

Every value is typed, scored, and traceable

A value is only useful if you can trust it. Alongside the typed result, Lirovo attaches a confidence score to each value so you can set your own threshold and gate anything below it.

Every value also links back to the exact second it came from, and whether it was spoken or shown on screen. So when the schema says the valuation is $14B, you can open the moment that says so and check it yourself. See how that works on the evidence page.

Explore the typed output

This is the Data view from the product, on the same run: the extracted fields laid out as queryable tables, each row carrying the source moment it was drawn from.

Mistral CEO Arthur Mensch

analysis ready

Mistral CEO Arthur Mensch

CNBC · 45:59

Timeline0:00 / 45:59

Arthur

Arjun

Events

0:0011:2922:5934:2945:59

Compute & sovereignty

AI economics

Semiconductors

Vibe & agents

Cybersecurity

AGI & physical world

ContradictionDecisionClaimclick to scrub

Speakers2

Arthur MenschCEO of Mistral

92%

Arjun KarpalCNBC host

video_facts_bronze·Mistral CEO Arthur Menschvideo_id=325gGv0eWV8hash=9f2a3c7d4e1b…job=job_c0296adaee7cmodel=gemini-3.1-flash-lite

fact_type·

Results17 rows · bronze · 1 row = 1 fact · validation: clean

#	type	subject	attribute	value	unit	conf	source
1	metric	Mistral AI	valuation	$13B+13,000,000,000 USD	USD	0.95	audio
2	metric	Mistral AI	valuation	$14B14,000,000,000 USD	USD	0.97	visual
3	metric	Mistral AI	data_center_investment	€4 billion4,000,000,000 EUR	EUR	0.96	audio
4	target	Mistral AI	compute_capacity	200MW → 1GW200,000,000 → 1,000,000,000 W	W	0.92	audio
5	range	US hyperscalers	ai_capex_per_year	$700–800B700,000,000,000 → 800,000,000,000 USD	USD	0.89	audio
6	metric	Europe	digital_deficit_vs_us_per_year	€250B250,000,000,000 EUR	EUR	0.88	audio
7	metric	AI inference market	token_market_size	~$5T5,000,000,000,000 USD	USD	0.84	audio
8	metric	Global manufacturing	addressable_market	$30T30,000,000,000,000 USD	USD	0.83	audio
9	metric	ASML	stake_in_mistral	~11%11 %	%	0.90	audio
10	entity	Mistral AI	organization.name	Mistral	—	0.99	audio
11	entity	Arthur Mensch	speaker.role	CEO of Mistral	—	0.98	both
12	entity	Vibe	product.name	Vibe	—	0.94	both
13	entity	ASML	organization.role	Lead investor (Series C)	—	0.91	audio
14	claim	Mistral AI	invests_in	data centers in France and Sweden	—	0.95	audio
15	claim	Europe	has_window	roughly two years to build sovereign compute	—	0.86	audio
16	claim	AI agents	disrupt	the per-seat SaaS model	—	0.82	audio
17	decision	Mistral AI	ownership	remain a private company	—	0.90	audio

Use it anywhere

The same typed result comes out of the harness wherever you run it: standalone in your terminal, as a plugin inside your agent harness, or over MCP, so it drops into your pipeline, your workflow, or your assistant without a manual step.

Terminal

run it standalone, typed JSON back

Agent plugin

drop it into your agent harness

MCP

call it from your assistant

sources/capture/graph/example/evidence

Turn a video into a record like this

Run a live extraction on a video of your own, or join the waitlist and run it on your own machine with your own inference.

Run a live extraction Join the waitlist