Spent the last few months with my co founders studying why so many crypto trading platforms eventually fail users, even when the UI looks polished and the PnL screenshots look impressive.
The deeper we went, the clearer the pattern became:
Most systems in this industry still rely on one or more of these:
- black box signal generation
- custodial fund control
- unverifiable performance
- static risk models
- delayed execution
- selective reporting of wins only
That architecture might work in a bull market, but structurally it breaks under volatility.
Especially after FTX, Celsius, etc and the wave of “AI trading” platforms that disappeared the moment conditions changed, it feels like the industry still hasn’t solved the trust layer properly.
So we approached the problem differently.
Instead of asking: “How do we maximise signal frequency?”
We started asking: “How do you build trading infrastructure where the architecture itself reduces trust assumptions?”
That led us into some interesting design decisions:
• Non custodial execution only Capital remains on the user side, either on-chain or on their own exchange account. No pooled deposits, no transfer of custody.
• Multi model consensus instead of single-model prediction Rather than one system forcing directional bias, we experimented with independent models specialising in signal generation, validation and risk positioning separately.
The interesting part wasn’t even accuracy initially, it was disagreement.
Some of the best risk filters came from situations where one model strongly disagreed with the others during unstable market conditions.
• Bayesian confidence weighting Not averaging outputs equally, but dynamically weighting confidence depending on current market regime and historical calibration.
• Full trade transparency One thing we noticed in crypto is almost everyone shows entries, very few show invalidations, losses, drawdowns or live execution history.
But without seeing losses, there’s no actual way to evaluate robustness.
So we became obsessed with exposing the entire lifecycle of a trade:
- reasoning
- confidence
- risk parameters
- execution
- outcome
- post-trade attribution
Ironically, showing losses publicly ended up increasing trust more than showing wins.
Another thing we learned:
The real edge in this market probably won’t come from “one magical AI model.”
It comes from infrastructure quality:
- execution architecture
- latency handling
- liquidity awareness
- dynamic sizing
- risk compression
- cross market interpretation
- human override systems
- capital preservation logic
That’s where the moat actually starts forming.
The closer we got to production environments, the more obvious it became that trading systems behave less like prediction engines and more like probabilistic operating systems under uncertainty.
Curious how other builders here think about this.
Especially people working on:
- agentic finance
- execution systems
- Hyperliquid tooling
- quantitative infrastructure
- AI consensus systems
- risk engines
- autonomous trading agents
Do you think the future belongs to:
- fully autonomous systems,
- human AI hybrid execution, or
- infrastructure layers that orchestrate multiple specialised agents together?
Feels like the industry is still very early in understanding what “AI-native financial infrastructure” actually looks like.