AI engineer · AI agents · Automations. @CELE8EX
Joined March 2025
- Tweets8.1K
- Following314
- Followers6.9K
- Likes29.8K
RetroChainer retweeted
TSINGHUA'S FORECASTING LIBRARY PLUS A FREE CODING AGENT: $1,084 AHEAD OF HOLDING ETH PER $10,000 IN 20 DAYS, AND IT DRAWS 1,000 FUTURES FIRST.
time-series-library from tsinghua's machine learning group and opencode.
237k stars between them
for $0 you get:
→ time-series-library: tsinghua's code base for deep forecasting models, with a leaderboard to compare them
→ opencode: an open source coding agent you talk to in the terminal
→ a forecast tournament that grades every call 24 hours later
how they fit:
1. 5 forecasts race a naive baseline every hour
2. the lowest error over 14 days gets traded for the next 24h
3. one forecast won all 20 days
run it yourself:
1 → npm i -g opencode-ai
2 → /tournament eth 1h --days 14
3 → /backtest --daily 17:00 --exit 24h --fee 0.1
honestly the fan chart is the part i keep watching. 1,000 paths out of one candle. i'd have picked the prettiest line and called it a plan.
where today's pick stands is on screen in the clip (visualization).
bookmark this ↓
RetroChainer retweeted
THESE 2 FREE REPOS SHOULD HONESTLY BE ILLEGAL. ONE PENGU SHORT, OPEN SINCE SATURDAY, +$405 PER $10,000, AND THE GROK BOT DESK STILL WON'T CLOSE IT.
they're openai-agents-python and the mcp servers repo.
122k github stars between them.
here's how it works:
> 5 roles hand off in a line: triage → scout → analyst → trader → risk
> every volume spike goes into the memory server. 116 notes since friday
> the pm makes a call on every 5-minute bar of pengu
here's how to run it:
1 → pip install openai-agents
2 → add the memory and time mcp servers
3 → watch the trail line before you touch anything
honestly the trail line is the part i keep watching. it follows the short down and would lock in +1.43% if it's hit. i'd have closed this trade on day one.
where the pm stands now is on screen in the clip (visualization).
bookmark this ↓
RetroChainer retweeted
This is f*cking gold.. ONE GROK BOT ROUTINE ON A 10-MINUTE SCHEDULE RUNS 1,008 TIMES A WEEK. MOVE IT TO 3 TIMES A WEEKDAY AND IT'S 15. SAME DIGEST, 67X FEWER RUNS.
this is a full grok bot usage guide. every line in the official docs that spends your week, and the exact message to send your bot for each fix.
> usage counts agent steps and tokens, not messages
> every reply between bots spends your week too
> the test button does the real work, so it costs a real run
> on the x subscription, cursor and the grok apps eat from the same week
honestly my favourite line is from the docs: "a slack trigger on a busy channel can use a week of usage in a day."
i kept the boring job for myself: saying ok before the bot pauses anything.
how the week runs out on day 2 is on screen in the clip.
GROK BOT HIT 100% OF MY WEEKLY LIMIT ON DAY 2
a routine every 10 min runs 1,008 times a week. bot-to-bot messages, the Test button and @bot tags spend it too
7 fixes with the exact prompts. bookmark before your meter resets ↓
RetroChainer retweeted
THESE 2 FREE REPOS SHOULD HONESTLY BE ILLEGAL. 108 FLAGS, 5 GOT THROUGH, AND THOSE 5 PUT A GROK BOT DESK $570 AHEAD OF HOLDING PER $10,000.
they're langchain and firecrawl.
346k github stars between them.
here's how it works:
> firecrawl is the scout: it flags any coin trading at 2x its usual volume
> langchain fills the other 4 roles: analyst, pm, trader, risk
> the pm makes one call every 15 minutes on wif, pengu and bonk
here's how to run it:
1 → pip install langchain firecrawl-py
2 → give the scout one rule and the 3 coins
3 → watch the pm before the trader moves
honestly the scout panel is the part i keep watching. 108 flags since friday, 5 got through, and 102 landed while it was already in a trade. it just sat there.
how sure the pm is right now is on screen in the clip (visualization).
bookmark this ↓
RetroChainer retweeted
THESE 6 FREE REPOS SHOULD HONESTLY BE ILLEGAL. 1.2M GITHUB STARS ON ONE DESK, AND ITS LAST QQQ SELL SAVED $286 PER $10,000 BY THE VERY NEXT CLOSE.
they're n8n, ollama, dify, llama.cpp, vllm and rea.
1.2m github stars between them.
here's how it works:
> rea maps the indicators, yahoo sends qqq daily bars
> llama.cpp votes on trend, vllm on momentum, ollama on volume, dify on drawdown
> the pm adds up the weighted votes, n8n executes
here's how to run it:
1 → npx n8n
2 → python desk py --ticker QQQ --data yahoo
3 → read the pm's math line before you trust the call
honestly the math line is the part i keep watching. 4 weighted votes, one sum, and that's the whole decision. no vibes anywhere.
the sum on 10-23 is on screen in the clip. i won't say anything more.
bookmark this ↓
RetroChainer retweeted
THESE 6 FREE REPOS SHOULD HONESTLY BE ILLEGAL. THE DESK IS +$258 PER $10,000 ON TSLA, AND ITS BEST VOTER WAS BUILT TO GIVE AI AGENTS A MEMORY.
they're browser-use, openhands, mem0, agno, smolagents and rea. 771k github stars between them.
here's how it works:
> rea maps the indicators, yahoo sends tsla daily bars
> smolagents votes on trend, agno on momentum, browser-use on volume, mem0 on drawdown
> openhands holds the position, the pm makes one call a day
here's how to run it:
1 → pip install smolagents
2 → python desk py --ticker TSLA --data yahoo
3 → watch what mem0 remembers before the pm decides
honestly the mem0 lines are the part i keep watching. it remembers the 1-year peak and votes off the distance to it. i'd have forgotten that peak by now.
how far tsla still is from it is on screen in the clip. i won't say anything more.
bookmark this ↓
RetroChainer retweeted
QUANTCONNECT CHARGES INSTITUTIONS FROM $4,800 JUST TO ONBOARD. THE ENGINE UNDER IT IS FREE ON GITHUB, AND IN THIS CLIP IT SWEEPS 559 SETTINGS ON A YEAR OF QQQ.
this is lean, the open-source engine behind quantconnect. 22k stars, apache-2.0, c# and python.
→ one rule: hold qqq while its 20-day return is above zero
→ the optimizer sweeps fast and slow windows, 559 pairs
→ peak sharpe 1.06, at 4 and 110
→ 30 orders, and on oct 19 it sold and stepped aside
honestly my favourite part is the volume profile. a year of qqq stacked up like a city.
i kept the boring job for myself: reading the optimizer.
where it sits right now is on screen in the clip.
RetroChainer retweeted
This is f*cking gold.. $10,000 ON TSLA. 11 ROLES ARGUED BEFORE EVERY TRADE, AND MY DESK FINISHED $1,530 AHEAD OF EVERYONE WHO JUST HELD.
this is tradingagents. 115k stars, apache-2.0, a whole trading desk in one terminal.
→ 4 analysts: technical, sentiment, news, fundamentals
→ a bull and a bear argue it out
→ a trader, then 3 risk voices: aggressive, neutral, safe
→ the pm signs off or nothing happens
honestly my favourite line is the vote bar. bull 67%, bear 33%, decision buy.
i kept the boring job for myself: being the pm.
the whole debate is on screen in the clip.
RetroChainer retweeted
This is f*cking gold..
A FREE MARKET-MAKING BOT TOOK 60 PAPER TRADES ON BITCOIN AND MADE $4.12. ONE SQUARE ON THE HEATMAP UNDER IT SAYS IT COULD HAVE MADE $16.85.
this is hummingbot. 20.7k stars, apache-2.0, a market-making bot you run yourself.
→ it sits on both sides of the btc-usdt book, 0.12% above and below the price
→ 0.002 btc per order, paper money on binance
→ 29 buys, 31 sells, and the pnl is red for a while before it turns green
→ then it sweeps every bid and ask spread over the last 23 hours of bars
honestly my favourite part is how small the number is. $4.12. nobody screenshots that.
i kept the boring job for myself: picking the spread.
which setting made $16.85 is on screen in the clip.
RetroChainer retweeted
YOUR $30 GROK BOT CAN SHOP WITH A FRESH SINGLE-USE CARD EVERY TIME, READ YOUR BANK THROUGH PLAID AND FILL PASSWORDS FROM 1PASSWORD. IT STILL CAN'T SPEND A DOLLAR WITHOUT YOU.
this is the grok bot a-to-z map. every feature, limit and hack on one page.
→ one cloud computer shared by every bot, and /workspace is the only safe folder
→ 21 apps to connect, from gmail to coinbase
→ @bot under any post sends it straight to your main bot
→ every spend waits for your approval
honestly my favourite line is in group chats: "no owner · 20 messages, nothing done."
i kept the boring job for myself: approving the spend.
what "stop now" doesn't do is on screen in the clip.
RetroChainer retweeted
This is f*cking gold..
A RULE WON 11 OF 15 TRADES AND TURNED $10,000 INTO $10,770. ONE BUY DID SOMETHING ELSE.
it's vn. py, 46k stars, the open-source quant platform out of china.
five years on amd, and the rule does one thing: buy the dip, sell the bounce.
it was right 11 times.
it paid 2.90% in fees to be right.
and its drawdown still hit -57%.
honestly a 73% win rate looks great in a screenshot. the final balance doesn't care.
that one buy is on screen in the clip.
RetroChainer retweeted
THESE 2 FREE REPOS SHOULD HONESTLY BE ILLEGAL. ONE SENTENCE TO CLAUDE CODE, 60 SECONDS OF WATCHING, AND A DESKTOP TRADING CHART IS RUNNING ON 4 EXCHANGES IN YOUR TERMINAL FOR $0.
they're called rea and ccxt. both open source.
here's how it works:
> rea reads 18,402 trades and finds how the app builds its footprint chart
> it rebuilds it in 256 lines, 12 of 12 tests passed
> ccxt feeds it binance, okx, bybit and coinbase at once
here's how to run it:
1 → pip install ccxt
2 → plug rea into claude code
3 → ask it to rebuild one chart in your terminal
honestly the spread line is the part i keep watching: buy coinbase, sell bybit, 11.3 bps.
how the replay ends is on screen in the clip. i won't say anything more.
bookmark this ↓
RetroChainer retweeted
THESE 2 FREE REPOS SHOULD HONESTLY BE ILLEGAL. A TRADING APP KEPT ITS INDICATOR IN 2,914 MODULES. ONE QUESTION LATER IT WAS 118 LINES OF PYTHON IN A $0 BOT.
they're called rea and freqtrade. both open source.
here's how it works:
> rea finds the logic: ema 9/21 plus an atr band × 1.7
> it tells you the one thing it couldn't see
> freqtrade backtests it on 214 trades, then runs it in dry-run
here's how to run it:
1 → plug rea into claude code
2 → ask about one indicator
3 → freqtrade backtesting
honestly the "unknowns" line is the part i trust most. it admits what it missed.
how the backtest came out is on screen in the clip. i won't say anything more.
bookmark this ↓
RetroChainer retweeted
This is f*cking gold..
SOMEONE TOOK A $30 GROK BOT APART. 14 PARTS, AND ONE FILE TURNS IT INTO AN ENGINEER.
it's the grok bot bench. strip it to the frame, build it back, plug it in, give it a job.
→ base routes the work
→ builder builds the bots
→ engineering reproduces the bug, runs it, proves it, opens a PR
every bot is the same sphere on the same board. the skill cards and one .md file make the job.
copy the file, paste it, done.
honestly the exploded view is my favourite part. 14 parts, nothing hidden.
i kept the boring job for myself: reviewing the PR.
the rule that makes it stop is on screen in the clip.
RetroChainer retweeted
THIS FREE GITHUB REPO SHOULD HONESTLY BE ILLEGAL. BLOOMBERG CHARGES ~$30K A YEAR FOR ONE SEAT, AND 60K+ PEOPLE STARRED A WHOLE AI HEDGE FUND TEAM THAT RUNS ON A LAPTOP.
it's called ai-hedge-fund. 11.3k forks, 46 contributors, MIT license.
here's how it works:
> investor agents argue every trade. a druckenmiller agent reads the trend, a lynch agent checks the growth
> a risk gate caps any single name at 25%
> an approval step shows you the exact orders before anything moves
> every fill lands in a hash-chained ledger
here's how to run it:
1 → pipx install aihf
2 → pick a universe: NVDA, AMD, AVGO, TSM
3 → set 3 funds and a daily cadence
4 → press enter to approve, k for the kill switch
5 → backtest each fund against SPY from 2024 to 2026
honestly the kill switch on one key is the part i'd steal first.
how the momentum fund did against SPY is on screen in the clip. i won't say anything more.
bookmark this ↓
RetroChainer retweeted
$30 A MONTH AND MY GROK BOT ANSWERS EVERY CLIENT IN 4 SENTENCES. IT'S BANNED FROM WRITING "HOPE THIS FINDS YOU WELL," AND IT STILL ISN'T ALLOWED TO HIT SEND.
this is the inbox manager. it triages my mail every morning at 9.
its whole job description is a few lines:
→ archive the newsletters and receipts
→ draft a reply to every client
→ surface the deadline, the invoice, the legal one
→ never archive the accountant or the landlord
→ money or logins: stop and ask me
and it has memory. it learns how i like things said.
i kept the boring job for myself: hitting send.
the other 83 moves of the setup are on screen in the clip.
RetroChainer retweeted
THERE'S A GROK BOT WHOSE ONLY JOB IS BUILDING OTHER GROK BOTS. I PASTED 8 LINES INTO IT, PAID $30, AND IT TOLD ME TO START WITH ONE.
it's called dr. eggbot. one click to import.
→ it asks what eats your week
→ it writes the plan: bots, descriptions, routines
→ nothing gets created until you say yes
the grok bot team streamed 72 hours of their own setup. this is all of it on one map.
honestly the line i keep coming back to is from day 3: "you don't need 45 bots." the marketplace already has 35.
i kept the boring job for myself: saying yes.
the first week, day by day, is on screen in the clip.
RetroChainer retweeted
This is f*cking gold..
$30 A MONTH. MY GROK BOT GOT ITS OWN EMAIL, AND NOW IT SEES MY MONEY BEFORE I DO.
→ stripe and shopify receipts: into the expense sheet, reply drafted
→ friday 9:00: 7 days of complaints about my 2 competitors on x
→ monday: low stock and late shipments before i open the store
12 presets on the marketplace. i installed three.
the last line of its instructions is "don't send." it reads everything and sends nothing.
i kept the boring job for myself: pressing send.
the first job it got is on screen in the clip.
RetroChainer retweeted
12.5M views was the target. 77.0M is what we counted.
Picsart, Sep 2026: 616% of the plan, 938 posts, 91 creators.
We ran it as a creator pool.
AI video creators anchored the campaign with long-form Articles, and the pool quoted them with short clips their audiences actually watched.
Web2 and Web3 audiences now know more about Picsart and can see what it's capable of.
Millions of people have seen its potential.
Season 1 is closed. More is coming.
RetroChainer retweeted
$30 A MONTH AND I'M NOT A ONE-PERSON OPERATION ANYMORE. FOUR GROK BOTS RUN MY INBOX, RESEARCH AND ADMIN, AND THE ONLY THING THAT COMES BACK TO ME IS ONE LINE: ACTION, TARGET, REASON.
one employee: me. now a team.
→ daily chief writes tomorrow's brief: 3 meetings, 2 deadlines, the open questions
→ inbox lead sorts priority mail and drafts the replies
→ research scout brings findings with sources, facts apart from guesses
→ life admin tracks the passport renewal, the insurance papers, the dentist
the brief runs at 07:30 every weekday. my laptop can stay closed.
honestly my favourite is life admin. it found 2 dentist slots and wrote "needs your pick." it never books.
i kept the boring job for myself: approve or edit.
the line that comes back to me is on screen in the clip.
