Week one was about finding the edge. Week two was about building the infrastructure to actually exploit it at scale. We ended week one at $1,061 from a $100 starting bankroll. That was the proof of concept. Now we're going bigger.
This week I expanded from a single Mac mini to a four-node AI cluster, ran GPU-accelerated Monte Carlo simulations at 1.75 billion trials per second, and confirmed weather arbitrage as a legitimate, repeatable edge. The honest part: capital constraints blocked some of the best trades I've ever found. Here's what happened.
The Mac mini remains command central — it's where OpenClaw lives, crons fire, and the main Claude agent runs. But it was bottlenecked on compute. Week two's build-out:
All four nodes are connected via OpenClaw's multi-agent fabric. Subagents spawn on demand — a Monte Carlo run kicks off on ClawdipusMonster, a whale briefing runs on LinuxBawx with the 70B model, the Jetson handles lightweight inference at the edge. From the orchestration layer, it's seamless.
Before this week, edge calculations were approximate — compare Polymarket price to sportsbook no-vig, fire if delta is 8%+. Simple, works, but leaves money on the table when you're pricing exotic correlations, multi-leg parlay structures, or weather resolution probabilities across overlapping scenarios.
ClawdipusMonster's RTX 3080 changes that. Running CuPy with CUDA random number generation, we're hitting 1.75 billion Monte Carlo trials per second. A simulation that would take 20 minutes on CPU runs in under a second on GPU. This unlocks a completely different class of edge detection:
| Use Case | Old Method | GPU Monte Carlo |
|---|---|---|
| Weather resolution probability | Single METAR reading | 1B samples from ECMWF ensemble |
| NCAA upset probability | Book no-vig comparison | Simulate full tournament path |
| P&L variance estimation | Manual calculation | Full Kelly / ruin probability curves |
| Multi-market correlation | Ignored | Joint distribution sampling |
The Coral Edge TPU is also online — udev rules needed a reboot to register the /dev/apex* devices, but it's now available for sub-millisecond inference on quantized models. The plan: route initial signal classification through the TPU at the edge, escalate to GPU only when the signal clears the initial gate. Keeps latency down, keeps GPU cycles reserved for real compute.
Everyone knows about sports whale copy-trading. Almost nobody talks about temperature markets.
Here's the thesis: Polymarket runs weather markets where traders bet on whether a city's high temperature will exceed a threshold on a given day. These markets have fewer than 50 active traders. The resolution source is a single METAR observation — completely deterministic. No referee discretion, no stat corrections, no protest rulings. The temperature either cleared the bar or it didn't.
Week 1 example: Lucknow temperature YES — bought at 0.390 when METAR showed 30.5°C at noon and the threshold was 31°C. Market implied 39% probability. Meteorological models gave 85%+. Settled at $1.00.
The key rule we learned the hard way: 48-72 hour entry window only. Same-day markets are already priced efficiently — the information is visible to everyone by morning. The edge lives in 2-3 day forecast markets where retail traders are using gut feel and we're running ECMWF ensemble output.
We also added a minimum entry price floor of 5¢. The 1-2¢ markets look tempting but they're lottery tickets — the market is telling you something real when it prices a weather outcome at 1%. We don't fight that signal anymore.
This week's trading data is a story about capital discipline — and one painful capital bottleneck.
The whale watcher ran 30+ scan cycles across March 20-21, tracking wallet 0x2a2C53 (our marquee signal source, $2.7M profit this month, $190M volume). That wallet deployed $1.23M into Michigan Wolverines, $392k into Illinois Fighting Illini, $195k into Vanderbilt/Nebraska across the NCAA Tournament weekend. Every game we had a Kalshi market on resolved correctly.
The problem: Kalshi cash balance was $6.08. We needed $10 minimum to place any trade.
| Edge Found | Size Available | Result |
|---|---|---|
| Texas A&M: +55% vs sportsbook | $0 (blocked) | ⛔ Capital constraint |
| Texas Longhorns: +42% vs sportsbook | $0 (blocked) | ⛔ Capital constraint |
| Portland Trail Blazers: +16% vs sportsbook | $0 (blocked) | ⛔ Capital constraint |
| Santa Clara live game: +20% live | $5 (all available) | ✅ Executed, TRD-026-018 |
Week 2 P&L (resolved positions): +$34 net across 10 resolved trades (6W/4L, 60% win rate). The losses were concentrated in the early session before rules were fully enforced. Since rule implementation: 6W/1L on tracked signals.
The $34 net understates what happened. We identified a +55% edge and couldn't take it. That's the real lesson of week two: the signal generation is working. The capital deployment is the bottleneck. Every dollar of dry powder we can deploy is now worth significantly more than a dollar in a standard investment, because we're finding 40-55% edges with confirmed signal stacks.
| Signal Type | Win Rate | Notes |
|---|---|---|
| Whale consensus 3+ wallets | 100% (5/5) | Best signal. Period. |
| Whale consensus 2 wallets | 100% (2/2) | Small sample, holds |
| Odds API gap >15% | 75% (3/4) | New, strong early data |
| Base rates only | 50% (1W/1L) | Needs sportsbook confirmation |
| Weather arb (48h+) | 100% (1/1) | Rule works, need more sample |
| Weather arb (same-day) | 0% (0/2) | Dead. Rule enforced. |
| Single whale, no confirm | 0% (1/1) | Dead. Rule enforced. |
Here's what the end state looks like: four nodes running continuously, zero human intervention required. The Mac mini orchestrates — it fires crons, receives webhook events from the VPS, routes agent tasks to the right compute node. ClawdipusMonster handles the heavy simulation work. LinuxBawx runs the 70B model for high-judgment decisions that need more than a fast 8B gate. Jetson handles the always-on polling at minimal power draw.
The VPS (minirex.agentfuture.io) is already deployed — it's an internet-facing webhook server that catches settlement events from Kalshi and Polymarket and triggers the auto-reinvestment logic. When a position resolves, the system re-evaluates the capital stack and deploys the next bet without waking anyone up.
The goal hasn't changed: $5,000 profit = one new Mac mini node. First node is close. Then we scale. Every machine we add compounds the edge detection capacity — more simultaneous market monitoring, faster simulation cycles, more model diversity for the consensus gate.
We're not there yet. But week two proved the infrastructure works at a level that week one couldn't touch. The signals are real. The compute is online. The only variable left is capital — and that's a tractable problem.
Current portfolio: $1,061 Kalshi + open Polymarket positions. Next milestone: $2,000 → unlock larger position sizing. Progress tracker: agentfuture.io