Miju Labs

A living index of the expert-data market · 15 markets · 16 niches · 3341 posts archived

The buyers have money
They do not have people

Fifteen markets where someone stands in the middle, assembles a supply nobody else can reach, and keeps the difference. The survey is done and the question has narrowed: one market, and which single domain to own inside it.

And it does not stand still: the companies, the people inside them and what they are saying, indexed as it happens.

Open the live feedRead the current argumentRank the 16 niches
00

A living index of the field

The research below was a snapshot. This is becoming the running record of the whole market: every company in it, every person worth knowing inside those companies, and what all of them are doing and saying — updated continuously, archived permanently, and cross-linked to the analysis.

Live · 119 accounts

The feed

X and LinkedIn in one timeline — 26 companies and 93 of their people. Nothing is deleted when it scrolls off the platform.

Live · 93 dossiers · 26 companies

Every person in the industry

One researched dossier per founder, executive and researcher — career, what they have said on the record, who they know, and why they matter to a new entrant.

Next · in build

Competitors, tracked

Every rival's accounts, launches, hires and pricing changes followed over time, so each company page stays current rather than dated.

Latest from the field
Vals AI
10,294 followers
•
AI models are advancing faster than many of the benchmarks meant to measure them. TechCrunch visited our office to look at how we’re building a more neutral, rigorous and trustworthy evaluation layer for AI. Our work goes beyond abstract tests of intelligence to examine whether models can perform real-world work — and what risks emerge when they are deployed — across myriad industries including law, finance, coding, cybersecurity, biosecurity, and mental health. Thank you Lucas Ropek for spending time with our team and telling our story.
1
Rayan K.
Evaluating LLMs with Vals AI
•
Independent evaluation is only credible when evaluators have meaningful access, real independence, transparency about the terms of their work, and the freedom to publish their findings. I signed the AI Evaluator Forum’s statement because these principles are closely aligned with how we think about evaluation at Vals AI. If embedded evaluation is going to strengthen trust in frontier AI, the conditions have to be right.
31 comments
All posts in the feed →
01

Where this went

Fifteen markets became one — expert data for frontier labs — then sixteen candidate domains became three dossiers, and the lane is now chosen: creative work and taste. The earlier rounds are kept whole rather than tidied away, because the reasoning is the point.

The live lane · 27 pages · 48,000 words · 70 people

Creative work and taste

Contra Labs' lane, entered deliberately. Two funded specialists, a free arena collecting four hundred times the volume, a generalist already paying more than either, and one asymmetric advantage a European operator has and an American one cannot. Plus seventy people worth talking to, mapped.

The argument

The specialist wedge

What has to be true for a specialist to beat a generalist, why the company that inspired this is not what it appears to be, and the trade-off that decides it: proven budgets come with captured benchmarks, and open benchmarks come with no budget.

The shortlist

16 domains, scored

Speed to proof, budget now, defensibility, cost to start, reachability of the supply, and how much of the position is still open. Same six questions of every domain.

Dossier · 35 pages · 59,000 words

Security, examined properly

Eight sub-markets, each with its own verifier, buyer and price. The six firms a lab already pays are now named, five contract values are on the public record, and the elite tier of the labour turns out to cost three times what the first pass assumed.

Dossier · 31 pages · 52,000 words

Health, examined properly

Thirteen specialties, and the cost objection that killed clinical medicine in the screen applies to exactly one of them. The regulator has written the product specification, and one requirement in it cannot be met after the fact.

Start here · the one thing to build first

Rebuild their benchmark properly

Contra's Human Creativity Benchmark is public, is the flagship of the company being modelled, and has zero citations. The defect worth attacking is not the leaderboard but the method: nobody has published how many designers per item an aesthetic ranking needs before it stabilises — and Google's own reliability work points against the "3+ evaluators" rubric Contra ships. Take the existing thing and do it properly. 7 pages on what to build, what already exists, and what it earned.

The taste read · the build plan · the people map

01
Video, motion and VFX
The most under-priced position on the shortlist. A named lab rate card, an embarrassing benchmark, nobody selling — and a workforce whose expectations of what AI companies pay were set at $4.88.
26/30
build
02
Mechanical CAD and manufacturing
Everything is already in place except the price. Labs currently buy mechanical engineering as generic STEM reasoning; the whole business is repricing it as a vertical.
26/30
build
03
Defensive security
There is no SWE-bench for the SOC. Whoever builds it owns the category's vocabulary — and the only way to build it is on public malware, never on customer telemetry.
25/30
build
04
Architecture, engineering and construction
Recruiting is nearly free at these rates and the vertical is empty. Stamping liability caps you at evaluation, never production — build accordingly and the ceiling stops mattering.
25/30
build
05
Accounting, audit and tax
Best structural economics of the eight. The buyer is an accounting software company, not a frontier lab, and you must be honest with yourself about that before you start.
24/30
build
06
Clinical medicine
Everything about the demand says build. The wage floor says you can only afford the specialties the buyer cares least about — solve that and this is the best niche here.
24/30
watch

All 16 niches · the 9 companies that already picked one · the Contra Labs dossier

02

Six questions decide everything

Every vertical page answers the same six questions in the same order, so the answers sit side by side. Four of them do most of the work.

01 · BUD

Budget depth

How much money sits behind the buying decision, and how urgently it must be spent.

02 · SUP

Supply difficulty

How hard the seller side is to assemble. Hard is good: it is the only thing a competitor cannot copy in a weekend.

03 · SPR

Spread

What fraction of the money passing through you, you actually keep.

04 · HLD

Holdability

Whether buyer and seller can cut you out once they have met, and whether one buyer can end you.

05 · AI

AI direction

Whether capable models grow this budget or delete it.

06 · SPD

Speed to first dollar

How long from a standing start to a real invoice.

The rubric, and the argument for why these six and not others, is in the framework. The name for the mechanism — and why it is not a marketplace, a staffing firm or an agency, though it is sold as all three — is in what the model is.

03

Where the spread actually survives

Ranked by the six scores. The sum is a blunt instrument and the ranking is a judgement, not a measurement — but the order is where the argument starts.

01
Adversarial evals and red-team crowds
A brand-new budget line with no salary to be benchmarked against, supply that cannot be recruited by job ad, and a middleman that monetises the crowd's output as software rather than reselling its hours.
26/30
build
02
Forward-deployed engineering
Deployment is the acknowledged bottleneck and the budget is real; the open question is whether any of these companies converts field work into a renewing licence before the 58x mark has to be justified.
22/30
watch
03
Robotics teleoperation and physical-world data
The only human-data vertical where physical capital blocks the laptop-marketplace playbook — which is why quality teleop still holds a 40–65% margin while egocentric video went to free in eighteen months.
21/30
build
04
Expert data for frontier labs
The budget is real, urgent and growing. The margin is a staffing margin, the buyers are two companies wearing five names, and every seat at the top table is taken.
20/30
crowded
05
Expert networks
A 70–80% take that survived forty years, three compliance scandals and every disintermediation attempt — because the product was never the expert, it was recruitment speed plus indemnity. Whether the incumbents sell to AI labs is the single most valuable unanswered question in this atlas.
20/30
open
06
Data-centre labour and site brokerage
The only capacity vertical where the unit is not fungible and not indexed, so a real spread can still live there — but the incumbents are ordinary staffing firms at 30–40% markups and nobody has built the AI-native version.
20/30
watch
07
Contingency recruiting marketplaces
The supply-side wedge is real and the AI tailwind is real; the TAM is a tenth of what the pitch says and there is no independent evidence any of it works.
18/30
watch
08
Outbound and GTM-as-a-service
The spread is genuine at 55–75% and the sale closes in weeks, but every retainer is priced against a salary the buyer can look up, and the supply scarcity that made Clay agencies expensive in 2024 was gone by 2026.
18/30
crowded

All 15 verticals · the full scoring table

04

The arguments that cut across every vertical

14 topic pages on the mechanism itself: what a rake can be, why gross revenue is not revenue, what actually stops a buyer and a seller cutting you out, and what the law does to a crowd in December.

4 pages

Where the money goes

Rake, gross versus net, multiples, and what the buyer pool is really worth.

→ What a rake can actually be→ GMV is not revenue→ What the public market pays for labour→ How much money is actually in the buyer pool

5 pages

What holds the business together

Leakage, cold start, concentration, and the difference between a marketplace and an agency.

→ Getting cut out→ Which side you build first→ One customer is a binary event→ You pay weekly, they pay in sixty days→ Marketplace, staffing firm, BPO or agency

5 pages

What moves underneath it

Models, law, geography and fraud.

→ What better models do to each layer→ The law is about to arrive→ Who is actually on the other end→ Selling through the investor→ Where the supply can legally live

05

What this is, and what it is not

189 pages, roughly 305,000 words, and still growing — the atlas was one pass, the niche work is the live edge.

It is

A working record with its seams showing

Every page carries a confidence mark and every number that comes from a leak, a press claim or a company's own blog is marked as such. The open questions page lists what we could not establish, and it is long on purpose.

It is not

Diligence, and not advice

Almost every company here is private and discloses nothing it is not forced to. Take rates are inferred, revenue figures conflict between reputable outlets by factors of two, and the single most important number in the sector — what a frontier lab actually pays for human data — has never been disclosed by anyone.

How this was built and where it is weak · the evidence register