Workforce datasets are static.
A snapshot. A list. A moment in time.
But companies are not static.
They grow.
They contract.
They shift role composition.
They reallocate talent before revenue changes show up.
So instead of building another database, I built a longitudinal company-year panel.
~2.5M normalized U.S. companies.
~387M company-year rows reconstructed from historical experience timelines.
Median 7 years of workforce history per company.
Not profiles.
Not contact records.
Company-year intelligence.
For each company and each year:
• Observed headcount
• Growth rate
• Role distribution shifts
• Structured entity normalization
The real asset isn’t volume.
It’s the ability to ask:
– When did this company actually start scaling?
– Did engineering grow before sales?
– How did workforce composition change pre-funding?
– Which segments show consistent multi-year expansion patterns?
Longitudinal structure turns raw records into signal.
Investors call it alternative data.
Strategists call it market intelligence.
AI teams call it training infrastructure.
I call it organizational time-series intelligence.
Building this in public.
#infrastructure #database #patternrecognition
the crowd count is known before you buy, so no lookahead. thats the right way to build it
the split holding in both halves of the week is the strongest part. 63 vs 25, then 54 vs 25 is not nothing
two things id still check tho. why those 202 devs? if they were picked because their coins got traded, the sample might lean toward survivors. and 24 trades in the losing group is thin, one more week could move that 25% a lot
we see something similar from the other side. in our data price and volume in the 100 to 200k zone barely separated runners from rugs. who bought and when tells you way more than what the chart did. a launch where everyone arrives in the same 10 seconds is usually the same few people in different wallets
run it forward for a week with zero changes. if 50%+ in 10 sec still loses, thats a real filter
we use it every day, but mostly not the way people expect
whats worked: AI as the builder. scanners on every launch, backtests, wallet checks, alerts to telegram, a full panel. stuff that used to need a dev team, built by 2 people with claude code. if youre starting with zero dev background thats honestly the best use rn
whats failed: AI as the picker. we tested whether attention on X predicts runners and it didnt, by the time a coin is getting talked about the move is mostly done. we also let an LLM call which coins would 3x and it said yes to zero of them and missed every single one that ran
so no, it doesnt find the plays for you. it scans everything while you sleep, tests your ideas in an hour instead of a month, and kills the bad ones fast. the 3am call is still yours, you just stare at 5 coins instead of 500
this is the most honest memecoin backtest ive seen on here in a while. most people post test 1 and stop
one thing id add to your point about "coins people bought": check whether the buys in that sample happened after minute 1. if yes its lookahead, because the sample already knows which coins survived long enough to get bought. same for gmgn trending, it only works as a filter if the coin is trending at the moment you'd enter
we found almost the same thing in our lab. a rule that bought young bases early passed on coins we already knew broke out, then failed on every base with no hindsight (0.90x). and the 100 to 200k zone was brutal across the board: 0.78x on average, 86% halved
your question 3 is the one most people skip tho. 18 trades is a coin flip with a story attached
Everyones racing to build fully automated trading bots.
We went the other way: the system does everything except press buy…
It watches thousands of coins, scores every setup against its history, ranks 16 models against a random baseline, and pings me when something lines up!
But the buy button stays mine. once im in, the machine takes over: stops, partial sells, trailing exits. no emotions, no "maybe it comes back"
Why not let it run alone? because most of what we test fails. the bot is great at the boring 90%: discipline, speed, never getting tired at 4am.
Its terrible at the 10% that matters most, reading whether a coin is actually alive.
I had a coin where the system said wait and i was interested. it went +68% in under an hour. other times my gut was wrong and the stop saved me…
So the split is simple. machine for the rules, human for the judgment, and the rules always win once the trade is on
my worst trades were never the machine's fault. they were the ones where i overrode my own stop.
totally normal imo. look back 70 years and every big tech wave had a brutal correction right in the middle of it
the dot com crash took the nasdaq down almost 80% and the internet still ate the world afterwards. railways, electricity, same story. the tech was real, the timing and the valuations werent
corrections like this shake out the tourists, not the trend. in a few years when we zoom out, the people who faded AI at this point will look like the ones who sold amazon in 2001
shit was brutal, sure. but it usually is right before the real run
@fxnction@opensea it is a real challenge. I didn't manage to get me out of the loop, but we are close. However, I really like the combination of the me and the agent.
Pterodactyl. NUREMBERG. WIGGLES. DOUG.
17 Nov 16:31
Sounds like a group chat, not a watchlist...
My system flagged all four in the last three days, when each was worth between 39k and 61k.
100 dollars in would have become 368, 361, 684 and 371 at their highest points.
Silly names. Serious moves 😛
I think we’re on the verge of the biggest opportunity in human history to achieve financial freedom and live on our own terms.
Most people have no idea how close it is.
Nobody puts "we stopped growing in 2021" on their website.
17 Nov 09:00
But the hiring history knows...
Which kinds of teams they added. Which ones they never did. The year new roles dried up.
I followed 196 companies from 2019 to 2026. About 1 in 3 were no longer independent by the end.
Deal makers buy the snapshot. The signal is in the movie 🎬
How about we forget about constantly building products for others and use the world class power of AI to make cash for ourselves instead of serving the ones already being over served…
Just a thought…we’re entering a new era where clients won’t be the asset to success.
100%, but theres a trap in it. almost everyone calls their losses variance and their wins skill
you only know a loss was variance if you wrote the plan down before the trade. entry, stop, why. if you followed it and it still lost, fine, thats variance. if you made it up as you went, it wasnt variance, it was just a bad trade with a good excuse
we do the same in our research lab: rules frozen before we look at results, every time. most of what we test fails, and thats what keeps the few winners honest
the win you got away with is the dangerous one tho. thats how overconfidence starts, and for me that was always the expensive part
fair warning, but "i dont care if im wrong, we can always long later" is the most honest line in the post. thats the whole game for chart calls rn: zero cost to the caller, full cost to whoever follows it
the level is real tho. 650 to 670B has been the floor all year, and everyone sees the same green box. when a level is that obvious it usually gets wicked first to clear the stops, and only a daily close below it with volume actually means something
and 473B is roughly 27% down from here. if that happens, alts dont fall 27%, they fall 50 to 80%. so if youre worried, sizing down beats calling the target
agree, but ppl are hyped about the wrong part
whats underhyped is what it lets a tiny team do. we're 2 people and with AI we ran pre registered backtests on our historical workforce data across 77k US companies, 2013 to 2019, in days not months
one thing that came out of it: companies that swap out half their top skills in three years have a much wider range of outcomes. for established companies with average growth, the odds of scaling doubled (4.2% vs 2.0%). a human analyst would never get through that many company histories
whats overhyped is thinking AI just knows the answer. it doesnt. it makes testing your ideas 100x cheaper, and most of them still fail. thats the part nobody gets rn
Agree on distribution, that's the whole game. I spent years in performance marketing, and the product is rarely what wins. The cheapest customer acquisition wins.
But be honest about what you'd be building. A normie friendly perps app is a machine for handing leverage to people who shouldn't have it. In our own testing, leverage and memecoin volatility don't mix: the stops get hit before the move ever comes.
The business model works because most users lose. The fees come from the losers.
And it's not empty ground. Several wallets already route perps to Hyperliquid behind a clean interface. The gap isn't the app. It's who owns the audience, and whether you can live with what your users are paying for.