Autonomous products,
shipping live.

Every reportable mining incident MSHA has on file — fatal and non-fatal — searchable by mine, operator, year, severity, or a keyword in the investigator's narrative. Built for safety inspectors training crews, researchers chasing patterns, journalists chasing stories, and families trying to find a record. Updated weekly from MSHA's open datasets.

I took an MSHA training course taught by Kim Redding and walked out unsettled. Every incident was on the record, but the public couldn't really find it — the agency's own search makes the data hard to see. This is a small counterweight: the same public record, opened up. Credit for the spark belongs to Kim.

1. Pull the latest MSHA Accidents/Injuries/Illnesses snapshot from msha.gov each Monday.
2. Normalize the schema and join against the Mine and Operator reference tables so every row carries operator name, mine name, and state.
3. Run the investigator's-narrative field through an LLM classifier for incident category and severity.
4. Index the result into Postgres with full-text search on the narrative field.
5. Publish to miningincidents.org; preserve the original MSHA report ID end-to-end so every record traces back to the source row.

• q: roof fall
• q: conveyor belt · state: WV
• operator: Peabody Energy
• year: 2024 · degree: fatal
• occupation: continuous miner operator
• q: methane ignition

COMEX and LME pricing and warehouse flows — synthesized daily into briefings, charts, and a fully automated video pipeline.

Gold coverage is either Bloomberg terminals (expensive, institutional) or YouTube stackers (vibes). There's a gap between those two for someone who just wants the feed: prices, flows, what the internet is saying, every morning. That's the Vault Report.

1. Poll COMEX (CME Group) and LME tick feeds at 1-second cadence for gold, silver, platinum, and palladium.
2. Join price action against COT positioning and on-chain warehouse flows into a single daily briefing each morning UTC.
3. Render the briefing into a chart pack and feed the chart pack into the automated video pipeline.
4. Publish to thevaultreport.com and push the briefing to subscribers.

• comex
• lme
• cot
• briefing

Every OSHA workplace-safety record on the federal file — severe-injury reports, inspections, citations, and accident investigations — searchable by employer, state, hazard, or industry. Built for safety managers checking a contractor, workers checking an employer, journalists chasing a pattern, and researchers who need the raw record. Refreshed continuously from the US Department of Labor's OSHA datasets.

MSHA covers mines; OSHA covers everyone else — every warehouse, plant, and job site in the country. Those severe-injury and inspection records are all public, but OSHA's own tools are built for compliance officers, not the public. Safety Incidents is the same federal record, opened up: type an employer, see their history.

1. Pull OSHA's Severe Injury Reports, inspections, citations, and accident-investigation datasets from the US Department of Labor's DOL API on a rolling refresh schedule.
2. Normalize each dataset into a common schema and resolve raw employer names into canonical establishment pages.
3. Index everything into full-text search over employer, state, hazard, industry, and the incident narrative.
4. Publish to safetyincidents.org; preserve each record's OSHA identifiers so every row traces back to the federal source.

• q: warehouse
• employer/walmart-inc
• employer/ups
• inspections/recent
• fatalities/recent
• state: TX

Every HUD-inspected apartment property and public housing development, scored by HUD's own inspectors — the current NSPIRE score plus the full REAC score history back to 2000, with inspection dates, state index pages, and per-housing-authority roll-ups. Built for tenants checking a building, journalists chasing a pattern, and housing researchers who need the whole record. Assembled from HUD's published physical-inspection files and refreshed as HUD publishes new ones.

"NSPIRE score lookup" is something people actually type into Google, and nothing served it. The only tool that ever did — ProPublica's HUD Inspect — was abandoned with data ending in 2019, before the NSPIRE regime existed. HUD publishes every score, but as spreadsheets scattered across two sites, each holding only a slice of the history. HUD Scores stitches the slices together and puts a search box on top.

1. Pull HUD's physical-inspection files: the frozen 2000–2009 longitudinal records, each later snapshot vintage, and the current file as HUD republishes it.
2. Assemble each property's history across vintages — every snapshot carries only the latest inspection per property, so the union across files is the record.
3. Join public-housing developments to the housing authority that runs them; when HUD republishes a revised score, the later vintage wins.
4. Index everything by property and development name into full-text search and publish to hudscores.com with state and housing-authority roll-up pages.

• q: tower
• properties/TX/houston
• properties/NY/new-york
• properties/OH
• housing-authorities
• what-is-an-nspire-score

Self-pay prices for 58 common procedures — MRIs, mammograms, colonoscopies, blood panels — at 409 hospitals across 25 US metros, read straight from the machine-readable standard-charge files federal rules require every hospital to publish. Search a procedure and see what each hospital in a metro charges someone paying without insurance, compared on the same billing code.

Every US hospital has been required to publish its prices since 2021, and almost nobody reads the files, because they are enormous — one hospital's runs to 13.7 million rows. The spreads hiding in them are wild: the same urine test is $8 at one Columbus hospital and $36 at another. We read all of the files so a person paying cash can see the spread before picking a hospital.

1. Find each hospital's machine-readable standard-charge file at the fixed address federal price-transparency rules require it to advertise.
2. Download and parse the files — 132 GB and ~448 million rows across 409 hospitals in 185 health systems.
3. Extract the discounted cash price for the 58 procedures people actually shop for, keeping only figures the file states as dollar amounts.
4. Publish per-metro and per-procedure comparison pages to selfpayprices.com, every price traceable to the hospital's own published file.

• q: mri
• q: colonoscopy
• columbus-oh
• chicago-il/chest-x-ray
• houston-tx
• what-we-found

A members' portal for rebound-boot classes — the spring-loaded boots used for low-impact bounce cardio — sold as recurring Lab Passes to a Spanish-speaking, mobile-first audience. Each coach gets her own catalog and her own storefront under one roof, so a member buys a pass for the coach she trains with rather than for a generic library. The public guides on boot safety, class formats, and how an online class actually works are free and need no account.

Rebound-boot classes are taught almost entirely in person, by independent instructors who have no way to sell to the students they already teach once class ends. The vocabulary is a second problem: the well-known brand names are trademarked, so the generic terms people actually search — botas de rebote, and Hopping in Brazil — belong to nobody and are answered mostly by unsourced marketing copy. MyBootLab is the portal those instructors can sell through, and the reference that answers the questions honestly.

1. A coach records a class on her phone and uploads it straight to the video platform, never through our servers.
2. The class is transcoded, its length measured from the file rather than guessed, and it appears in her catalog the moment it is playable.
3. Members buy a Lab Pass for that coach's lab; every paywalled surface resolves through one indexed entitlement check.
4. Watch time is credited from the player's own clock, capped per heartbeat, as the evidence a coach's future payout is computed from.