Most people's AI agents are brilliant and forgetful.
Every new chat starts from zero. You re-explain your project, your preferences, your context β again. The intelligence is there. The memory isn't.
So I gave my whole AI team one shared brain.
One Obsidian vault. Every agent reads from it and writes back to it. When one agent learns something, they all know it. Knowledge compounds instead of evaporating at the end of a session.
The setup is simpler than it sounds: β One folder of plain-text markdown notes β Agents save durable lessons as they work β not throwaway chat β Notes link together, so any bit of context is one hop away β A new agent inherits the entire brain on day one
The result: no re-explaining, no amnesia, and a team that gets sharper every session instead of resetting to zero.
If you're running AI agents and starting cold every time β this is the fix. Steal it.
#buildinpublic #AIagents #SecondBrain #Obsidian
Quite an interesting pause with my ai-agent Roman to stop and learn the bond markets.
I think this is extremely overlooked by retail, and at this current time, it's probably more important than every for retail traders to learn and understand the relationships of the bond markets.
I'm going to ask Roman to write up a plain-English article later to better help retail understand.
Most "diversified" portfolios are one bet in a trench coat. Carry, short volatility, merger arbitrage, relative-value β they look different on paper. They all go up smoothly for years. Then they all crash in the same month. Here's why π
In calm markets they look independent because money rotates between them. But in a real crash money doesn't rotate β it stampedes into cash. Forced sellers dump what they CAN, not what they want β often your safe haven.
Correlations snap to 1.
The calm-time numbers were lying.
One strategy is structurally different: trend-following. Long-volatility by design β it pays when markets move hard in any direction. Positive in 2008. Positive in 2020. Not because it predicted the crash. Because it rode whatever happened.
I spent today reading ~135 quant papers with AI agents to map this properly. The lesson: real diversification means different mechanisms, not just different tickers. Most "multi-strategy" funds stack the same trade six times and call it diversification.
A few months ago I traded like most retail do:
β’ Gut feel
β’ Hot names
β’ "Technical analysis" with no actual statistics behind it β just hindsight dressed up as a signal
I've since stopped trading entirely to properly learn quant finance.
The first thing that broke: I stopped trying to predict.
Nobody knows if a ticker goes up or down tomorrow β it's speculation with confidence laid on top. What retail (me included) always neglects is the thing that actually decides the outcome: risk, survival, consistency.
A reality check that landed hard β most hedge funds, with armies of quants and infrastructure, return somewhere in the 5β15% a year range. Retail traders are adamant they'll beat that. Almost nobody does.
The eye-opener: the way the pros actually build portfolios was never the secret. Affording the quant team to run it was. With AI, you can now do it yourself.
Here's the shiftπ‘
β using Β£1,000 as the worked example:
What retail actually does:
Picks 3 assets, usually the ones that have been going up.
Puts 50% in the hottest name because it "feels right." Watches it tick every day.
Has no plan for a crash.
That's not a portfolio. It's a bet with extra steps.
What an engineered portfolio does β a basket chosen by role, not hunch:
Growth engine (world equities): the long-run compounder
Ballast (long government bonds): rises when equities panic
Real assets (gold, broad commodities): hold when paper burns
Crisis alpha (a trend sleeve that can go short): historically made money in 2008
Cash: the dial β controls how much risk you're actually running
The insight most people miss:
Β£1 of shares is not the same risk as Β£1 of bonds.
Equities are ~3β4Γ more volatile.
So split Β£1,000 fifty-fifty by money and you haven't split the risk β you've handed ~80% of it to the equity side.
Engineering means sizing so no single sleeve can wreck you.
Rebalanced on a rule, not a feeling:
Every week the engine trims what grew and tops up what shrank.
No "should I buy the dip?"
No timing.
Removes the emotion β removes you, mostly. Which is the point.
In a crash (2008):
Buy-and-hold equities: β50 to β57%
Classic 60/40: ~β30 to β35%
All-weather / risk-parity: ~β22%
A real trend sleeve: gained that year
And here's why that gap matters far more than it looks: drawdowns are asymmetric.
Lose 50% and you don't need 50% back to recover β you need 100%.
A half-wipeout takes a doubling just to break even.
Lose 22% and you only need ~28% back.
Avoiding the deep hole is worth vastly more than chasing the high.
The book bruises; it doesn't break.
That asymmetry is the whole point β survive, and compounding does the rest.
The boring truth: day to day this is dull. Nothing to watch, no dopamine. The edge is the structure and the discipline, not the screen time.
Retail builds to win when things go up. An engineered portfolio is built to survive when they don't. Survival is what compounds.
#Trading #Finance #Ai
Most trading strategies that beat the market in backtests are just luck.
Test 10 versions, keep the best β you've kept the luckiest, not the best. It falls apart on new data.
Today my AI quant system's auditor said this about my own risk engine: can't rule out luck. One real crisis in the data isn't enough.
So it refused to sign it off.
That's the win.
Today I shipped "the honesty machine" β the part of my AI quant system whose entire job is to say NO.
Before any strategy risks real money it has to pass seven checks in order: is there a real economic reason this should work? Is the track record skill or just luck (adjusted for how many things you tried)? Is it overfit? Would it have survived 2008, 2020, 2022?
Only a GO clears all seven. And a human still signs off β the AI never trades.
My own adversarial auditor then tried to break it. It built a fake strategy that looked honest β it just disguised a 2,000-trial search as a single honest bet, which switched off the luck-correction entirely. Got a false GO. My machine had been fooled.
My own agent caught it. Fixed structurally. Then a second reviewer found the safety maths weren't even running in the full test suite. Also fixed.
A solo builder has no one to disagree with them. So I built the disagreement in.
187 checks pass. Nothing is live.
#BuildInPublic #AI #Quant #QuantFinance
I built 'Reggie' β a custom data mining system that replaced $1,000+/month in third-party APIs from #Apify, #Outscraper, #RapidAPI, and #Firecrawl.
It now runs for just $12/month on a VPS, with no limits and equal or better quality.
All built with Claude and a team of AI engineering agents.Most things are possible with AI β if a third-party tool exists, chances are you can replicate it yourself.
Happy to share the repo for those interested.
I'm building an AI agent team for investing research. Lead quant is an AI I call Roman. Right now we're at the most unglamorous part: just wiring up data feeds.
But I built the foundations β including an adversarial auditor agent β before the first line of real code.
Day one of wiring data, the auditor flagged a corrupt vendor price: a fake +1,783% one-day move. Not a return β corrupted data. All 18 automated tests had sailed straight past it.
If the foundation hadn't been there first, that number would have silently poisoned everything built on top.
Build the boring stuff before the exciting stuff. Every time.