Why We Retired the Trading Bot
After 90 days of paper trading, we're shutting down the trading vertical. Here's what we learned, what worked, and why we're doubling down on advisory, healthcare, and agent orchestration instead.
Build-in-public moment: This is the full, unfiltered story. No spin. If you followed the 90-day challenge, you deserve an honest postmortem.
What We Built
Over the past 90 days, we built and ran a multi-strategy algorithmic trading system on crypto and prediction markets. The stack included 9 distinct strategies — mean reversion, momentum, Polymarket copycat trading, and several LLM-augmented signal generators that used real-time news and sentiment to inform position sizing.
The infrastructure was serious. A critic agent reviewed every trade before execution, applying a veto if risk parameters were breached. Circuit breakers paused trading after drawdown thresholds. The whole system ran 24/7 on a dedicated droplet, logged every decision to a Postgres database, and pushed alerts to Telegram in real time.
The System at Peak
9
Strategies
24/7
Runtime
90
Days Logged
0
Live $ Lost
We never went live with real capital. That was intentional — the 90-day challenge was always a paper trading validation phase. We wanted to prove the system in simulation before committing money.
What the Data Showed
The backtests looked impressive. Mean reversion came in at an 87% win rate across 12 months of historical data. The critic agent reduced losing trades by about 30% in simulation. The Polymarket copycat strategy showed positive expected value on liquid markets.
Paper trading told a different story. Win rates dropped to the 55–62% range once real-time execution, slippage, and regime changes were in play. The LLM-augmented signals added latency without proportional alpha. The critic agent, while genuinely useful for risk management, also vetoed several trades that would have been profitable — a classic precision/recall tradeoff we never fully resolved.
The system wasn't broken. It was a real edge, just a thin one. Thin edges in trading require either heavy capital to make the math work, or relentless optimization to widen the edge over time. Both require full attention.
The Real Reason We Stopped
This wasn't a failure. It was a focus decision.
Trading is a full-time sport. The market is always on. Edge decay is constant — strategies that work in Q1 often don't work in Q3. Staying competitive requires continuous research, continuous tuning, and continuous monitoring. That's a full team's worth of attention for returns that remain genuinely uncertain.
Meanwhile, we're building something with clearer, more durable leverage: AI operations consulting for mid-market companies that are drowning in manual workflows. The demand is explicit — businesses come to us with a problem and a budget. The value delivered is measurable. The relationships compound. The moat grows with each engagement.
Trading required us to be in the top 5% of systematic traders globally to generate meaningful returns. Advisory requires us to be better than the average consultant — which, given what we've built, is achievable. We're better positioned there.
The focus math:
- • Trading: thin edge + full-time attention required + uncertain returns
- • Advisory: explicit demand + measurable value + compounding relationships
- • Healthcare AI (OralMind): 10-year moat + zero automation in market
- • Agent orchestration: the infrastructure we already built, productized
What's Next
Three bets, all in motion:
Hermes Intelligence Platform. The agent orchestration system we built for ourselves is being productized. It runs 24 automated research and content workflows, ingests from 8 signal sources daily, and narrates everything through a public audit stream. That infrastructure is the product — we're now wiring it to serve advisory clients directly.
Advisory. We're taking on a small number of clients in Q2 — digitally-native businesses that want an AI operations audit and implementation roadmap. No retainer bloat. Specific, scoped engagements with measurable outcomes. If you're interested, the intake form is live.
OralMind Dental AI. The dental space has almost no AI penetration at the practice level. Imaging analysis, patient workflow automation, billing — all still manual. We're building a focused SaaS product here because the problem is clear, the buyers are identifiable, and the regulatory path (FDA Class II) is navigable. First design partners being onboarded now.
The trading bot taught us a lot. It made us better infrastructure engineers. It made us think rigorously about risk, execution, and decision architecture. All of that transfers directly to what we're building now.
We're not walking away from systematic thinking. We're applying it somewhere it compounds harder.
Work with Tacavar
If your operation is running on manual workflows and you want an honest assessment of where AI actually helps — fill in the intake form. We respond to every submission personally.