01 / 07

tidy.work

Scroll to explore

Software built for the work you'd rather not do

Case Study — 01 / 05

Sony Music

Record labels live or die on signing unsigned artists before competing labels do. We built Disruptor a daily feed of the best new ones — no TikTok scrolling required.

Case Study — 02 / 05

MedLegal

25× cheaper. 12× faster. Personal-injury firms usually pay a team of paralegals to read 1,000-page medical records and analyze every hospital bill — now, MedLegal does it in minutes for a little more than a dollar.

Case Study — 03 / 05

Coefficient Labs

Contact-form outreach converts better than email, but filling forms by hand doesn't scale. So we built a robot that does it — and drove hundreds of thousands in closed sales.

Case Study — 04 / 05

Morehands Maids

A cleaning company couldn't keep up with reviewing customer calls by hand. We built them an overnight system that grades every call and tells managers exactly what to coach.

Case Study — 05 / 05

Bluesky Energy

A solar company's field crews were snapping hundreds of receipts a month — and a bookkeeper was re-typing every one into QuickBooks. We replaced the bookkeeper with a robot.

tidy

Case Study — 01 / 05 · Disruptor Records, Sony Music

Sony Music

A record label lives or dies on signing great artists before anyone else does — and that means scrolling TikTok all day. We built a tool that does the scrolling.

Dailycurated feed of unsigned artists who posted in the last 24 hours

100%of surfaced artists are real signing candidates, not noise

$0.003to classify a single video — pennies a day at scale

The problem

The music industry is old school. A&R managers — the people whose job is to discover new artists — spend hours every day scrolling TikTok by hand looking for someone worth signing. Most of what they find is noise: reposts, fan clips, artists already signed to another label, or aspiring artists who aren't ready yet.

Whoever finds the next real artist first gets to sign them. Everyone else loses the deal.

What we built

A daily feed of the best unsigned artists who posted in the last 24 hours. Each video is automatically watched and graded — diamond (real, talented, worth a meeting), rough (interesting, watch them), or trash (skip). Diamonds are then cross-referenced against Spotify to confirm the artist actually exists, isn't already signed to a major, and is reachable. The result is a short, ranked list every morning, plus an email digest. No scrolling required.

Beyond discovery, the same pipeline doubles as a trend monitor — surfacing which kinds of clips, songs, and styles are actually performing this week.

Process note

The hard part isn't classifying videos — it's confirming the artist is real and available to sign. We chain several independent signals (profile metadata, link-in-bio, public catalog data) so by the time a scout opens the morning digest, "can we actually sign this person?" is already answered. That's the difference between a tool people use and one they ignore.

How it's built

Multi-model AI orchestrationVideo classificationCatalog cross-referenceDaily digest dashboardGoogle Cloud

Case Study — 02 / 05 · Personal-injury litigation

MedLegal

Personal-injury law firms pay paralegals to comb through massive medical records and analyze every hospital bill by hand. We do it in minutes, for a dollar.

25×cheaper than the manual workflow

12×faster, end to end

<5 minto analyze a case that used to take an hour

The problem

When somebody gets injured and sues, their lawyer has to prove how much was reasonably spent on medical care. The evidence is a 1,000-page hospital record — and somewhere inside it, scattered between scans, doctor's notes, and intake forms, are 30 to 100 pages of actual billing. A paralegal has to find those pages, read every line item, check whether the prices are reasonable, screenshot the proof, and assemble a courtroom-ready report.

MedLegal was doing this with a dedicated team. It cost roughly $25 a case, took about an hour each, and a single missed pre-incident charge was courtroom liability.

What we built

Drop a case file in, get a finished report back. The system walks the entire record, pulls out only the billing pages, identifies which hospital each one came from, extracts every charge, checks it against a reference table of fair pricing, and flags anything suspicious — including charges that pre-date the injury and shouldn't be included at all. It produces the same courtroom-ready document the paralegal team used to write by hand, complete with screenshots of the original billing pages as evidence.

It runs on the same shared drive the firm was already using, so there's nothing new to learn.

Process note

The reason this works is that the report still looks like a paralegal made it — same screenshots, same structure, same evidence trail. Lawyers don't want a black-box AI summary; they want the original page, circled. Build for the workflow that already gets used in court, not the one that's most convenient to automate.

How it's built

Multi-model AI orchestrationDocument understandingPage-region extractionSharePoint integrationReference-pricing engine

Case Study — 03 / 05 · B2B lead generation

Coefficient Labs

Filling out contact forms is one of the best ways to get a meeting with a company — better than cold email, almost as good as paid ads. The catch: doing it by hand takes forever. So we taught a robot to do it instead.

$100K+in closed-won sales driven by the funnel

100,000sof qualified companies contacted

Every steptracked from first touch to revenue

The problem

Most outbound sales is either cold email (cheap, low conversion, often ignored) or paid ads (expensive, no intent signal). Contact-form outreach is the sweet spot — you're reaching people who already invited inquiries — and it converts dramatically better than email. But filling out forms by hand is brutally slow. One salesperson can do maybe one or two an hour. That math doesn't work when you want to talk to tens of thousands of companies.

What we built

A system that visits qualified company websites, finds their contact form, fills it out with a personalized message, and submits it — at scale. Every site gets analyzed first to figure out what kind of form it has, so the system knows how to handle it correctly. Submissions get tracked end-to-end through a dashboard, so the team can see which industries and message variants are actually producing meetings and closed deals.

It's been responsible for hundreds of thousands of dollars in closed-won revenue.

Process note

The trick isn't filling forms — it's knowing what kind of form you're looking at before you touch it. The system fingerprints each site first, picks the right strategy, and only attempts the submission when it knows it will succeed. That's how we get reliability at scale instead of a 30% failure rate and a banned IP.

How it's built

Headless browser orchestrationSite fingerprintingAI-personalized messagingEnd-to-end attribution dashboardCRM integration

Case Study — 04 / 05 · Customer-service coaching

Morehands Maids

A growing cleaning company couldn't keep up with reviewing customer calls by hand. So we built an overnight system that listens to every call, grades it, and tells managers exactly what to coach.

Every calllistened to, scored, and coached — automatically

Monday-readycoaching dashboard each morning, no manual review

27distinct coaching categories distilled from years of past notes

The problem

Morehands is a residential cleaning company. Customer service happens entirely on the phone — bookings, complaints, rescheduling, refunds — and the quality of those calls is what keeps customers around. The problem was visibility: the owner was pulling call transcripts out of their call-tracking software one at a time, reading them, and writing coaching notes by hand. It didn't scale. Coaching ended up reactive — feedback only when something blew up, never when a pattern was forming.

What we built

Every customer call, every night, gets automatically transcribed, summarized, and scored against the company's own coaching playbook. Managers walk in Monday morning to a dashboard that shows them which agents need attention, which kinds of calls are going badly, and what specifically to coach — with the original audio one click away. The owner can update the coaching rules anytime, and the system retroactively re-grades using the new guidance.

Process note

The dashboard is only useful if the numbers reconcile. We made sure that when a manager filters to "complaints last week," the count matches what they'd get if they counted by hand — no silently dropped messages, no fuzzy aggregation. Operations leaders stop trusting AI dashboards the moment the totals stop adding up.

How it's built

Call ingestion & transcriptionMulti-model AI coachingEditable coaching playbookManager dashboardGoogle Cloud

Case Study — 05 / 05 · Renewable-energy contractor

Bluesky Energy

A solar company was processing hundreds of receipts a month, snapped on iPhones across dozens of job sites, in every format imaginable — and one bookkeeper was re-typing every one into QuickBooks. We replaced the bookkeeper with a robot.

Zeromanual receipt entry — field photo → QuickBooks, untouched

10+vendor receipt formats handled, including blurry scans

Hoursof bookkeeper time saved every week

The problem

Bluesky installs solar — which means dozens of active job sites at any moment, crews buying parts and supplies all day, and a pile of receipts at the end of every week. Field staff submitted them through a basic Google Form; from there, a single bookkeeper sat down and manually re-typed every receipt into QuickBooks, one at a time, assigning each to the correct job and cost category. The receipts arrived in every imaginable format — Home Depot thermal prints, smudged scans, iPhone photos taken inside a truck — and the bookkeeper had to make sense of all of them. Errors were common (wrong project, wrong cost code), and the rework ate into project margins.

What we built

Now field crews just snap a photo. The system reads the receipt automatically — figuring out the vendor, the amount, the date, and which job and cost category it belongs to — and posts the expense to QuickBooks without anyone touching it. When the system isn't confident about a field, it flags it for a quick review. The bookkeeper went from a daily queue of re-typing to a status notification.

Process note

The old workflow didn't get torn out — it got upgraded. Crews who wanted to keep using the original form still can; the new path layers on top of it. That's how you ship operational automation without forcing everyone to learn a new tool on day one.

How it's built

Receipt vision extractionChart-of-accounts validationQuickBooks integrationField-friendly capture flow