Mavrick @mavrickfitz.bsky.social · 20/07/2026Yep. I initially tried Google Form Parser for pure OCR table extraction, but the api pricing would have bankrupted me at scale. The LLM can also usually handle the custom spreadsheets scans with arbitrary layouts that some campaigns throw in, which feels like a magic trick when it works. 000
Mavrick @mavrickfitz.bsky.social · 08/07/2026The main route is actually an image quality score that Document AI OCR emits, with a 0.75 threshold. That usually catches the handwritten or badly scanned pages. Failing that, 15 or more corrections gets it flagged for proofreading. And dense spreadsheets turn off page batching. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026I've experimented, and dropping the the trad OCR increases the error rate, while bigger models don't improve results enough to be worth the cost at scale. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026Both. OCR first, then the LLM uses that as the primary source, fixes the mangled formatting, and has an image of the page to check if something looks off. It starts with gemini flash lite and upgrades to a Gemini flash with thinking tokens if it has to make too many corrections using the page image. 110
Mavrick @mavrickfitz.bsky.social · 08/07/2026Next steps: I'm going to expand this over the next few months with a backfill of previous election cycles, and then outward to other Tennessee counties. But more than anything I'm just curious what other possibilities are out there waiting for someone to find them. 000
Mavrick @mavrickfitz.bsky.social · 08/07/2026Not saying this is the end of professional programming. I'm absolutely certain that a professional developer using these tools could develop something better and more efficiently. But I don't see any professional developers volunteering to do this for my niche concern. 110
Mavrick @mavrickfitz.bsky.social · 08/07/2026"Campaign finance database for one specific county" is obviously a niche project, but if the difficulty gate and time commitment is so low that it's now feasible for a solo part-time amateur to build, I think there are things like this in every field and interest that are suddenly doable. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026Also, every time this codebase has threatened to exceed my grasp, I've been able to develop or refine automated processes for version control, code review, monitoring, and so on. It's cognitive off-loading, but the systems I'm offloading to have grown with the project. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026I can't stress enough that I wouldn't be able to do any of this without AI. I know it's associated with a whole flood of vaporware and crappy products, but I'm really excited to see what other kinds of projects are now possible by people with domain knowledge but no technical background. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026This isn't getting into some of the other non-trivial systems like compiling individual donation records into distinct entities, filtering small >$200 donor records out of the final product, daily updates, Datasette for export, or the whole astro frontend. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026Both sites make it possible to do things, like showing every candidate that a particular PAC has given to, or compare campaign totals between candidates at a glance, which are straight up impossible through the official portals. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026By contrast, pulling excel sheets out of a twenty-year old state website that errors out if you breathe on it wrong, and then figuring out their unwritten formatting rules, was easy-breezy. Everything is penny-perfect there. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026Basically, the system reads the form perfectly like 80% of the time. Another ~10% are fixable errors, ~5% reconcile to within $100, and ~5% have larger discrepancies no matter how hard I try, either due to filer math or being unreadable. There's a reason every page links the original pdf. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026This required evolving a whole system of checks, routing based on page type and difficulty, automated attempts to fix common errors, llm review, and tooling for faster human review. And then hours of human review squinting at bad handwriting and shoddy arithmetic. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026Naturally, the first thing I did (OCR on PDF filings) wound up being the hardest. A combo of trad OCR and an LLM with detailed instructions is near perfect on the standard forms when the responses are typed. But half of the forms had handwriting, and many campaigns attached their own custom tables. 200
Mavrick @mavrickfitz.bsky.social · 08/07/2026I have exactly zero experience with creating websites or automated data scraping, so it's been a wild ride. These aren't the prettiest or most sophisticated websites in the world, but there were very few equivalents for this at the state level in Tennessee, and nothing for local elections. 100
Mavrick @mavrickfitz.bsky.social · 08/07/2026Sharing this in case anyone is interested: Over the past 5 months I've been using Claude Code to develop a project that scrapes campaign finance filings from scanned pdfs (for Memphis) and excel files (TN state) and turns them into useable, searchable websites. whofundsmemphis.org whofundstn.orgwhofundsmemphis.orgWho Funds MemphisCampaign finance disclosures for every candidate in the 2026 Shelby County election — every dollar, every donor, searchable and free. 220
Mavrick @mavrickfitz.bsky.social · 26/06/2026This would have happened without Trump. The Census Bureau has a dual mandate to publish useful data while preserving privacy. Over the last decade, they've chosen to solve this by increasingly disregarding the "publish useful data" half of the mandate for small geographies. 020
Mavrick @mavrickfitz.bsky.social · 29/08/2025Going through stages of shock and bewilderment learning that in the UK, "Build to Rent" apartment buildings are apparently a controversial new 21st century development???? 021
Mavrick @mavrickfitz.bsky.social · 05/06/2025Post the craziest argument against building more housing that you've ever seen. I'll go first. 030
Mavrick @mavrickfitz.bsky.social · 25/04/2025This is of a piece with RFK Jr wrecking HHS. Complete abandonment of data expertise in favor of vibes, leading to decisions we know are going to kill people. 120
Mavrick @mavrickfitz.bsky.social · 25/04/2025Glad to see vehicular cyclists are having a normal one. 020
Mavrick @mavrickfitz.bsky.social · 24/04/2025Planning folks, please consider donating (if you're able) to this GoFundMe for the medical bills of a student in the first year of her Urban Planning grad program who was seriously injured in the mass shooting at Florida State. gofund.me/1ca8559cgofund.meDonate to Support Madison Askin's Healing Journey, organized by Dennis SmithOn Thursday, April 17, there was a mass shooting incident at Florida State Univer… Dennis Smith needs your support for Support Madison Askin's Healing Journey 000
Mavrick @mavrickfitz.bsky.social · 21/03/2025Hello from Memphis, one of the cheapest cities in the country. We also need a lot more housing production, especially in the city core where there are few options between crumbling 80 year old buildings and a handful of new "luxury" apartments. 010
Reposted by MavrickJohn Leavitt @leavittalone.bsky.social · 16/03/2025Always funny how big companies aren’t really in the industry they’re supposed to be , McDonalds is a real estate company, GE is mostly insurance, Amazon is a cloud compute company that outsources delivery to the USPS, Volkswagen sold more sausages then cars last year etc 1018533
Mavrick @mavrickfitz.bsky.social · 14/03/2025This is pretty awesome as a way to get temporarily rural bridge replacements on the ground as fast as possible. 010
Mavrick @mavrickfitz.bsky.social · 02/03/2025I'm sorry, when did the abundance movement go from "we can't enact a green new deal if we bury every bike lane and solar farm in red tape" to "logging endangered trees in the PNW is good, actually"? 030
Mavrick @mavrickfitz.bsky.social · 01/03/2025Over/under on Trump selling Taiwan for 100 leased pandas and the rights to build Trump Tower Hong Kong? 030
Mavrick @mavrickfitz.bsky.social · 27/02/2025Saw this shared on TikTok, the comments were all things like "I don't think there's that much urban housing in Maine" or "I didn't know we had sheep in Arizona" 🤦 4821
Reposted by MavrickKathleen Romig @kathleenromig.bsky.social · 26/02/2025I worked with US Digital Service at SSA. They were gems & we were all lucky to have them. They spearheaded SSA's award-winning website redesign. They were working to get simplified applications online for Social Security survivors & SSI. They were *actually* making gov more efficient. 213941
Mavrick @mavrickfitz.bsky.social · 24/02/2025Does anyone know where the belief that third places are "supposed" to be free/non-commercial came from? I keep seeing this on social media whenever coffeeshops, pubs, or breweries are brought up as places for in-person community. 030
Mavrick @mavrickfitz.bsky.social · 23/02/2025Good instinct, but those are state level laws. Federally preempting then might be worthwhile, but would not be quick or easy. 010
Reposted by MavrickAlfred Twu @alfredtwu.com · 20/02/2025In many ways it's fitting that the showdown between monarchy and democracy ends up being over NYC congestion pricing. Cities are where the power of kings were first challenged, and few things are more fundamental to government than control of public space. The cameras are staying on! 191717423
Mavrick @mavrickfitz.bsky.social · 17/02/2025I feel bad for not knowing, but is that normal in other (e.g. parliamentary) democracies or a product of weak mid-century American parties? 000
Mavrick @mavrickfitz.bsky.social · 16/02/2025Saw the logo first, was briefly concerned you had made a very questionable career move 😅 120
Reposted by MavrickKea Wilson @keawilson.bsky.social · 14/02/2025🚨: This got super buried in all the memos that followed Sean Duffy's confirmation, but on February 18th, USDOT is going to recommend that Congress rescind funding for anything the administration deems too "woke." Read: probably everything you care about. And it's time to call your reps.🧵usa.streetsblog.orgSec. Duffy Moves to Rescind Billions for 'Woke' Transportation on Feb. 18 — So Advocates Must Speak Up Now — Streetsblog USAThe U.S. Transportation Secretary has promised to call on Congress to slash vast funding for climate and DEIA. 64025
Reposted by MavrickKate Lowe (she/hers) @kateontransport.bsky.social · 13/02/2025This starter pack (from @aninehartmann.bsky.social & @cathytuttle.bsky.social) includes women and non-binary people bsky.app/starter-pack... This starter pack (from @chanface.bsky.social) includes a mix of gender and racial identities bsky.app/starter-pack... 23212
Reposted by MavrickTony Dunn @tony873004.bsky.social · 10/02/2025Here are 70 clones of #asteroid 2024 YR4 that do hit Earth, highlighting the impact risk corridor. There are some big cities along that line: #Bogota, #Lagos, #Mumbai. 1015556
Reposted by MavrickWilliam B. Fuckley @opinionhaver.bsky.social · 11/02/2025This is sort of the skeleton key to understanding a lot of shit: there 100% is a non-trivial amount of inefficiency and waste in the federal bureaucracy, and most of it is a *direct result* of ‘reforms’ that are meant to assuage the people who complain about “waste and inefficiency in govt” 403359705
Reposted by MavrickM. Nolan Gray 🥑 @mnolangray.bsky.social · 10/02/2025This amount of state capacity is unthinkable in Los Angeles. 946559
Mavrick @mavrickfitz.bsky.social · 10/02/2025Guarantee that like 40% of the public can't connect this to Trump's tariff threats. 021
Mavrick @mavrickfitz.bsky.social · 08/02/2025There's an infinite money glitch here, if we only have the courage bsky.app/profile/mavr... 020
Mavrick @mavrickfitz.bsky.social · 08/02/2025We have marriage rate data in Tennessee. Sevier County (Gatlinburg) is 9x higher than the county with the second highest rate. 100
Reposted by MavrickSean Skunk-Cabbage (Chou-Puant) ⚓ @moldybasil.bsky.social · 06/02/2025Walking? Woke 1173
Mavrick @mavrickfitz.bsky.social · 05/02/2025Love this. Also, I didn't know until recently that these photos of cuneiform tablets are practically life-size, with most tablets being smaller than the palm of your hand and packed with tiny writing. 1142
Mavrick @mavrickfitz.bsky.social · 05/02/2025We all appreciate you and every other civil servant refusing to comply. 030