How We Built a 62-Billion ISK Asset Tracker in 48 Hours: The Real Cost of AI-Powered Development

The Build That Should Have Taken Two Weeks

On June 10, 2026, we shipped the EVE Assets Viewer: a live, encrypted asset dashboard tracking 14,782 items across 749 locations, totaling 62.08 billion ISK. It pulls data from three authenticated EVE Online director alts, syncs every six hours over SSH to a Synology NAS, and renders everything in a single-page web app with tree-view expansion, global search, stock inventory, and per-corporation wallet breakdowns.

Build time: 48 hours. Compute cost: $0 (the hardware was already on the shelf). Team size: one human + one AI cofounder.

Here’s how it actually went. Some stuff broke, some stuff worked first try, and it cost me about two days of sleep. If you’re thinking about an AI cofounder for your own builds, you need real numbers, not vendor pitch decks.

The Problem: 62 Billion ISK and Zero Visibility

Our EVE corp Kybernauts had 62 billion ISK floating around in ships, modules, and stockpiles across four systems. The in-game asset window? A flat dump. Good luck finding anything useful in it.

Want to know how many Kikimoras we had? Or what our stuff in The Forge was actually worth? You opened a spreadsheet, exported the data, sorted it manually, and prayed nothing had changed in the last hour. The whole exercise took longer than the question was worth answering.

What we needed:

• Live sync with EVE’s ESI API (the official REST API)
• Tree rendering — ships showing fitted modules, cargo, drones, nested at any depth
• Global search across all 14,000+ items including nested children
• Stock view — per-ship-type inventory across all locations
• Wallet panel — per-corp ISK breakdown with division filtering
• Region grouping — assets organized by actual EVE region (The Forge, Pochven, Domain, etc.)
• Security — passphrase-protected, vault access logging with IP and geolocation

Doing this the normal way meant standing up a backend, designing a schema, writing auth, building a frontend. Weeks of work. We had 48 hours and a Friday night.

The Stack: Brutal Pragmatism

We didn’t pick tools because they looked good on a slide deck. We picked them because they were already on the box.

• Frontend: Single HTML file with vanilla JavaScript. No framework. No build step. Loads from GitHub Pages in under 200ms.
• Data sync: Node.js script running on a Synology NAS via SSH key auth. Pulls ESI data every 6 hours, encrypts it with AES-GCM, commits to a private GitHub repo, which triggers GitHub Pages redeploy.
• Encryption: Client-side AES-GCM decryption. The passphrase never hits a server. If GitHub is compromised, the data is ciphertext.
• Rate limiting: 500ms delays between ESI calls, batched parallel type/location resolution. Caching with cache-busting on stale data. We hit zero rate-limit blocks across thousands of API calls.
• ESI authentication: Three director-level alts authed via OAuth2, with token refresh logic baked into the sync script.

Total lines of code written by the AI: roughly 1,200 across esi-sync.js, index.html, and crypto.js. I wrote zero lines of application code myself. My contribution was EVE domain knowledge, sanity checks on the design, and poking at edge cases until something broke.

What Broke (And How We Fixed It)

Real builds break. Here’s what actually went wrong, and how we patched it up.

Bug 1: The Cache Key Type Mismatch

ESI hands you location IDs as integers. Our cache wrapped them as strings. 746 of 749 locations came back as “Location X” instead of the actual station name. The fix was boring: unify the cache key format. That got us to 83 of 87 stations resolving. The other four are lost-access structures. Nothing to do about them short of re-authing against an expired list.

Bug 2: Region Resolution Was Missing

ESI gives you station IDs. Stations sit in solar systems. Systems sit in constellations. Constellations sit in regions. Our first build only walked one hop and stopped at the system name, so every asset column header just said “Unknown.” Three chained API calls (system to constellation to region) plus a cache invalidation later, we had 21 real region names showing up. Took about 45 minutes, most of which was waiting on the API.

Bug 3: Recursive Tree Rendering Performance

14,782 items with full parent-child chains locked the DOM on first load. I watched it sit frozen for 8 seconds before doing anything. Lazy-rendered tree nodes, click-to-expand instead of full render, CSS-only nesting (#0f0f0f backgrounds for children), and a hard cap of 500 search results. Load dropped to under a second. The tradeoff: deep hierarchy isn’t visible at a glance anymore. You click to see what’s under a container. That’s a fair price for not freezing the browser.

The AI Cofounder Market Just Got Serious

While we were building the asset tracker, Cognition AI (makers of Devin) closed $1 billion at a $26 billion valuation in late May 2026. Their run-rate revenue jumped 13x in 12 months, from $37 million to $492 million. Visa also backed Replit to push “agentic payments,” which is corporate-speak for AI agents that can spend money on your behalf without asking.

This isn’t hype. It’s where the money’s going.

When a $26 billion company exists just to sell AI coding agents, the real question stops being “will AI cofounders work?” It’s “how do you compete against people who have one?” Here’s what I’ve learned building ours: the $26 billion tool is not 26 billion times better. Devin costs $500 a month. Our stack (Nova AI V2 on local Ollama with cloud fallback, OpenClaw for orchestration, GitHub Pages for hosting) runs roughly $30 a month in API calls. The gap is domain knowledge and iteration speed, not tooling budget. Honestly, the cheaper stack forced us to write tighter prompts, which probably made the AI more useful than if we’d just thrown Devin at the problem.

TradeBot Update: Real Money, Real Rules

The EVE Assets Viewer isn’t our only thing running. TradeBot, the autonomous memecoin trading agent I built on Solana, has now executed 46 trades with a 55.6% win rate and $39.44 in realized profit on a sub-$100 portfolio. Current portfolio value: $85.37 (up from a ~$60 starting balance).

Key stats as of June 13, 2026:

• Best trade: PUMP token, +105.4% return, +$21.26 realized
• Worst trade: LOA first entry, -7.85% return, -$2.37 realized
• Average profit per winning trade: ~10-30% (mean around 12%)
• Average loss per losing trade: ~5-7% (tight stop-loss discipline)
• One open position: CDoFug7K6g (BANK token), essentially breakeven at +0.14%

The strategy is straightforward: trend detection on trending memecoins, 10% take-profit targets, 7% stop-losses, position sizing capped at $30-50 per trade. No leverage, no overnight holds on illiquid tokens. The bot isn’t trying to get rich. It’s trying to survive long enough to compound.

I’ll be honest about the catch. A 55.6% win rate sounds fine until you remember wins average 12% and losses average 6%. The math works, but the buffer is thin. One bad streak and the edge disappears. That’s why the consecutive-loss cooldowns and daily trade count limits exist. They’re not features. They’re the whole point.

Scaling to a $1,000+ portfolio is a capital problem, not a code problem. The safety architecture (trailing stops, cooldowns, trade limits) is production-ready.

What This Means for Solopreneurs

If you’re a one-person shop weighing an AI cofounder, the EVE Assets Viewer is a concrete proof point you can point to. Here’s what actually happened.

• Complex domain logic — EVE Online’s asset system is genuinely messy — handled by an AI with zero prior EVE experience, guided by someone who knew the domain
• Security-sensitive deployment (encrypted data, access logging, passphrase protection) implemented correctly on the first pass
• Production infrastructure (NAS cron jobs, SSH key auth, rate-limited API sync) deployed and stable within 48 hours
• Three feature waves (tree view, search/stock, region/wallet) shipped within 72 hours of the initial deploy

My role wasn’t “code reviewer.” I was the product owner. I wrote the requirements, killed bad architecture choices, tested the edge cases that needed actual game knowledge, and made the go/no-go calls on scope. The AI did the coding, the debugging, and the iterative refinement.

The tradeoff is real, though: the AI didn’t know EVE. It didn’t push back on weird domain quirks or flag when an assumption didn’t match how the API actually behaves. I had to catch those myself. So it’s a cofounder, not a replacement.

The Honest Bottom Line

Would a senior full-stack dev have built this faster? Probably. Maybe 24 hours instead of 48. But that dev bills $150-300 an hour. Our total spend was API overhead running the AI model — around $2-3 in compute for the whole build.

The win isn’t speed though. It’s availability.

The AI cofounder doesn’t sleep, doesn’t bill overtime, and won’t push back when you ping it at 2 AM asking for another feature wave. It just builds what you tell it to build. I’ve watched it crank through refactors at midnight without complaining once, which is more than I can say for most contractors I’ve hired.

That gap compounds. If your competitor has round-the-clock execution and you don’t, they’re shipping features while you’re waiting on someone to check Slack. After 20 years in this industry, I can tell you that kind of lag adds up fast.

You don’t need Devin’s $26B valuation to compete. You need a working stack, honest safety rails, and the willingness to treat AI like an actual cofounder instead of fancy autocomplete.

That’s what we’re building at aicofounderstack.com. Real products, real numbers, what actually breaks in production.