Daniel Kalski

founder @Pre_Reason | not 30 |
building the data layer that sits between raw numbers and agent decisions // [email protected]

[Daniel Kalski @dankalski]

I left my career to build something I believe in.

Not another data API -> A context layer for the era where AI makes the financial decisions.

Raw numbers tell an agent what happened.

Context tells it what matters.

I tested the difference. 1,116 decisions. +15.97pp.

[Daniel Kalski @dankalski]

Good question; in a way, it's both.

Conditional because the two heads have to be back-to-back. The 25% from 0.5 * 0.5 answers "two heads anywhere in 2 fixed flips" which is a different question.

Then theoretical because 6 is the expected value across many trials/scenarios, aka the mean of the distribution.

In practice you might hit it in 2 flips or run over 20 flips before back-to-back heads land. Rare, but not impossible.

Definitely don't bet money with these answers 😂

[Daniel Kalski @dankalski]

Showing the product working beats describing it.

Visual simplicity >> Worded paragraphs.

I took this truth into my landing page.

> For example: Two terminal blocks side by side.
>> One shows the agent struggling to stitch raw data. >> The other shows it getting a clean briefing in 89ms.

No deep tech. Pure outcome.

Sell the the feeling of using the service. Not everyone has the time to sit through a deep dive.

That demo builds curiosity, and curiosity is what converts visitors. They see the agent doing the thing and want to know what else it can pull. Technical details, integrations, pricing all get read after the curiosity hooks them in.

Pages that lead with "we use advanced AI" lose the read entirely. Instead, show what the user GETS from investing their time and trust into your product.

[Daniel Kalski @dankalski]

I appreciate the clarity on peer comparisons and the bit-more outside care, especially when KTOS's value talk is in the randomness of high-valuation acceptance (by the current crowd) and on volatile geo-matters, aka when/how such a war moves forward.

And agreed, that what be your strategy with Claude is what's unique to your approach logic. To be seen in a few weeks/months on what that strategy returns back with.

[Daniel Kalski @dankalski]

@theaiportfolios btw on what I'm building → https://t.co/etMST6qakY

[Daniel Kalski @dankalski]

Re-reading these posts on how the model "justifies" the buy, yet same pattern keeps repeating:

Company context is well-defined. Wider outlooks are weak at best...

That's where the trade gets in trouble:

> FY2027 DoD budget treated as a sure thing. Election year + deficit talks make that political.

> 85x forward P/E with no peer comparison. AVAV, KRAS, TXT defense names sitting where on this scale?

> The "market treated CCA loss as a verdict, it's not" line is the LLM's own take. No data tied to it.

> Multi-quarter thesis but only May 6 mentioned. What's the catalyst path in between?

KTOS numbers are tight. The wider stuff, macro, politics, sector comps, is where these LLM reasonings usually go thin, almost as tunnel-visioning.

Same part that decides if the trade pays off.

Where's that data coming from in setups like this?

Macro reads, sector comps, political risk, packaged for the model in plain language, is exactly what I've been building.

[Daniel Kalski @dankalski]

Personally used most of these ideas as I got really into Claude Code in the summer last year.

Maybe 10 of the 100 listed actually move the needle. The other 90 cover the same 3 problems in different ways.

What works for me:

> Memory/Context tools. claude-mem ended my morning architecture re-explainers.
> Custom skills in CLAUDE.md. They compose with the codebase you're working on.
> MCP servers wrapping your own data. Custom ones compound, generic ones rot.

What hasn't:

> '100 subagents' collections. 80% never get used; too generic for your specific problems. Build the 3-5 you need.
> Other people's slash commands. They fit other people's flows. Write your own.

The 3 universal problems across these 100:

context loss between sessions, drift on long tasks, weak codebase-aware retrieval.

Good repos solve one. The rest cover the same ground.

[Daniel Kalski @dankalski]

For those wanting the juicy bits from the video:

> 22% of LSE daily volume = microsecond "sniping races". 5-10 microsecond windows, fired once per minute per stock.

> $5B/year hidden tax on global equity investors. LPs widen spreads to defend, investors pay in worse fills.

> The infra arms race went physical. HFTs built straight-line microwave tower networks Chicago to NYC, DC to NYC, to beat fiber-optic cables by milliseconds. Zero fundamental value to the market.

>> Why exchanges won't fix it: they make $1B+/year selling co-location and proprietary data feeds to HFTs. They monetize the broken design directly.

The fix is Frequent Batch Auctions (FBAs).

Trades clear in discrete intervals instead of continuously. Speed game dies; competition shifts from fastest to best price.

Flow Trading is the deeper move.

Instead of "buy 100 shares" you say "buy 1 share/second for 100 seconds", or "buy A and sell B only if their prices diverge".

Portfolio trades become native and one-leg execution risk goes away.

The part most crypto builders skip: TradFi latency arbitrage and DeFi MEV are the same structural flaw. The mempool makes crypto's version fully visible.

[Daniel Kalski @dankalski]

There will always be a need for 'the human touch' in anything marketing/visuals.

Lead sweepers, strategy templates, branding schemas... all can be built in a day, sure. I've shipped that myself.

What moves the bottom line (steady follower growth, viewer-to-user-to-paying-user funnel, becoming someone people look up to) needs what makes you YOU.

No markdown file holds that, despite the AI life we're being pushed into.

[Daniel Kalski @dankalski]

@Axel_bitblaze69 For how that work has been progressing: https://t.co/qDvyDrPbIk

[Daniel Kalski @dankalski]

The thing most people miss -> workflows are only as good as the data feeding them.

Doesn't matter how clean the automation is if the inputs are stale or generic web scrapes.

Half the repos on these lists assume the data layer is solved. It rarely is.

That's the slice I'm working on; turning fragmented APIs into briefings the LLM can read without making stuff up.

[Daniel Kalski @dankalski]

In most cases, I've seen people default to the websearch tactic with LLMs. Takes a few minutes, then the AI makes a judgement call on what it 'can' find.

The trick here most don't realize... the LLM comes back with what it could find, not what it should have found.

Doesn't take the time to realize what metrics are needed (NVDA - look at semi. industry health, foreign policy, global liquidity/rates), what's the overall view of that asset compared to XYZ; in relation to the global's ABC.

Best case; a few articles are found, a search here and there in 10-K docs, and a side piece written days ago.

You can get next-level with things by connecting to all the individual APIs. One for stocks, another for equity fillings, and more in macro/FRED/FX/Rates/etc...

But that's still half the work.

APIs feed raw numbers; the LLM still has to be told what those numbers mean in plain English.

From what I've experienced, LLMs 'reasons' a sentence way better than a CSV dump.

That's the slice I'm working on; turning all those APIs into briefings the LLM can read and move right into reasoning instead of 'guessing'.

Day to day testing my conviction here. Difficult, but as what's best said, it's worth a punt.

[Daniel Kalski @dankalski]

@AlphaCartell For those wanting to see what improved LLM inputs does: https://t.co/qDvyDrPbIk

[Daniel Kalski @dankalski]

The layer most of these don't address is the input layer (live, grounded, structured data fed to the agent before it decides).

Tools execute, models reason, but P&L follows the quality of input the first place.

Most of these stacks assume the data is clean. In practice, agents trip on stale price feeds, generic web scrapes, hallucinated context, and unstructured macro data.

Been building exactly that for months, ready for use.

[Daniel Kalski @dankalski]

Understanding context in financial markets is a hard problem.

Raw price prints, on-chain data, macro releases, FX moves... the model can take in all of it and still miss the actual message.

So I built the layer that synthesizes inputs into briefings the agent can read directly.

Interesting to see the theory I set on LLMs transforming into stone cold fact.

Linguistic input > Numerical spreadsheets

[Daniel Kalski @dankalski]

For too long I lived in fear of putting my work out there.

My mind, my way of solving problems, all of it.

In an age that demands attention/proof of work, my resume wasn't getting me anywhere.

> Scared of picking the wrong idea? just do it.
> Scared the demo breaks live? just do it.
> Scared the launch falls apart? just do it.
> Scared of bad feedback? just read it, just fix it.

Seeing the results is the only honest data point you get; for me, for whoever's watching...

So I keep shipping.

[Daniel Kalski @dankalski]

@TTrimoreau A never ending flywheel.

Product to have content to show,
build audience from what you show,
to show more you need better product results,
with better results/product, you spin up new content,

and so forth.

[Daniel Kalski @dankalski]

Bullish on the thesis so much that I sacrificed a weekend to ship x402 into my service.

But then you pull up the vol. chart though...

Agent economy looks like a dead memecoin 😅 peak in the Fall 2025, six-month bleed, $49M across every facilitator combined.

Can't stay forever down like this. The curve needs to ramp like a fresh equity IPO, day in day out.

[Daniel Kalski @dankalski]

And you say all that, but then the vol. chart shows the truth...

I'm all in on agents acting on their own accord, paying creator-set budgets, shopping APIs autonomously, etc.

A weekend was sacrificed to have x402 live on my service. The endgame holds.

Still, the x402 chart has the volume shape of a dead memecoin 😅

If this is to be the payment rail of the agent economy, volume would be ramping like a fresh equity on day one.

Let's get to work.

[Daniel Kalski @dankalski]

@tom_doerr Backtest evidence on input-layer impact, briefings vs no-briefings: https://t.co/GDddBuyNEK

[Daniel Kalski @dankalski]

88/100 on crypto-skill-bench tests how well the agent executes (routing, safety, UX).

It doesn't test whether the trade decision was right.

"AI autopilot for perps" fed on trending-tokens scrapes and stale on-chain data is auto P&L destruction.

Skills are downstream of context quality.

A trading agent making perp calls without grounded macro, on-chain, and cross-asset context will execute confidently wrong.