Build logs & deep-dives
Blog
Trading post-mortems, AI automation write-ups, and infrastructure build logs from Gordon Technology — what worked and what didn't.
Filter by Product
Three more runs of two AI agents in a box: the test set was the whole story
Runs 5 to 7 of the Claude plus Grok sandbox built a tariff classifier, a sanctions screener and a contract clause finder. One run missed every gate and said so, one scored its sealed set 22 times, one scored it once and the number held. Then we gave their ideas a GPU and a protocol.
Two AI agents, one sandbox, four runs: what made the third one worth reading
We put Claude and Grok in a locked-down VM four times and changed one thing per run. Two runs built tidy CRUD apps. One built a real job-shop scheduler and missed its own gate by 0.008 without moving it. Here is what each change did.
Our status page's first red marks were the monitor's own DNS
Five two-minute failures on October 8 sit in the status page history. Four were HTTP 200s that arrived 150 ms too late because the monitor's Wi-Fi Pi-hole was dropping queries. What we fixed, what we left in.
Gordon Lab Weekly — 4 experiments, every verdict published (2026-10-05)
4 experiments across 4 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 6 experiments, every verdict published (2026-09-25)
6 experiments across 6 research threads — the honest weekly roundup, losing runs included.
Write once, post to 13 platforms — no API keys, no sign-up, free
Crossfire: one post → 13 platform-perfect variants with pre-filled compose links. No API keys, no accounts to connect, free.
Which crypto coins can a simple ML model actually predict? (fresh run)
'ONDO' leads on auc: +0.542
Gordon Lab Weekly — 5 experiments, every verdict published (2026-08-26)
5 experiments across 5 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 10 experiments, every verdict published (2026-07-31)
10 experiments across 8 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 7 experiments, every verdict published (2026-09-21)
7 experiments across 6 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 14 experiments, every verdict published (2026-09-14)
14 experiments across 5 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 9 experiments, every verdict published (2026-09-07)
9 experiments across 6 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 5 experiments, every verdict published (2026-08-24)
5 experiments across 5 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 4 experiments, every verdict published (2026-08-16)
4 experiments across 4 research threads — the honest weekly roundup, losing runs included.
I Read X's Open-Source Algorithm. A Shared Link Is Worth 40 Likes.
xAI published the For You feed's actual ranking weights. Copy-link share = 20.0, a like = 0.5, a report = -234. Here's what the real numbers say about getting reach.
Gordon Lab Weekly — 9 experiments, every verdict published (2026-08-10)
9 experiments across 9 research threads — the honest weekly roundup, losing runs included.
My best trading strategy showed +62.8%. Then I tested it honestly.
The one profitable lead in my whole stack turned out to be drawdown insurance, not alpha — and optimising it made it strictly worse.
Gordon Lab Weekly — 49 experiments, every verdict published (2026-07-27)
49 experiments across 7 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 3 experiments, every verdict published (2026-07-17)
3 experiments across 3 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 21 experiments, every verdict published (2026-07-13)
21 experiments across 6 research threads — the honest weekly roundup, losing runs included.
Gordon Lab Weekly — 31 experiments, every verdict published (2026-07-05)
31 experiments across 8 research threads — the honest weekly roundup, losing runs included.
Are calm or wild coins easier for a model to predict?
'low-vol' leads on mean_auc: +0.509
I Funded a Solana Memecoin Sniper Bot. It Went 4 and 82.
I funded a Solana memecoin sniper with real SOL and let an XGBoost model pick every trade. 86 trades, 4 wins, 82 losses. The honest autopsy: execution was fine — there's just no edge at entry. Real numbers, including the one trade that actually worked.
A 0.70-AUC Crypto Model That Lost Money on Every Coin (AUC Is Not P&L)
I trained per-coin XGBoost models on 5-minute crypto bars. They hit ~0.70 out-of-sample AUC across 16 coins, walk-forward validated — genuinely predictive. Then I traded the signal and every single config lost money. Why a 'good' ML model isn't a profitable one: AUC measures direction, not magnitude.
Should a crypto grid bot re-anchor upward with price?
'trail3' wins: total_pnl +5542 vs 'static' +1858 (+3684, +198%)
Which crypto coins can a simple ML model actually predict?
'ARB' leads on auc: +0.532
Does crypto direction get more predictable at longer horizons?
'15m' leads on mean_auc: +0.526
Build Notes: A Week of Local-First Smart-Home Research
A research-driven recap of last week — what we tested, what we learned, and the gotchas that changed our plans while designing a fully local, no-cloud smart home around Home Assistant.
The Solana Bug That Failed 74% of My Sniper's Buys (and Lied About Why)
If your Solana transactions fail with Token-2022 "Invalid Mint" (0x2) or "simulation failed" on brand-new accounts, it's probably your preflightCommitment. The RPC default (finalized) lags the tip ~13-30s, so it can't see a seconds-old mint. The smoking gun, the cause, and the one-line fix.
YouTube Shut Down Our Channel. We Adapt and Overcome.
YouTube removed our channel under its scams policy and denied our appeal. Here's our honest take — what we do, what we don't, and how we're adapting and moving forward.
Front-Loading a Crypto Grid Bot: More Wins, Less Money
I tried to beat my bot's uniform 3% grid by front-loading the rungs to catch more dips. The aggressive versions hit an 81% win rate — and earned about a third less money. A backtest post-mortem on why win rate is a vanity metric for grid bots.
Optimization: Scaling Your Bot Profitably
Scale from 1% to 5%+ position sizing, add multiple strategies, and optimize for consistent 6-10% monthly gains.
Backtesting Crypto Grid Bots: The 3% Step That Worked (And What the Backtest Couldn't See)
I backtested grid bots before risking real money — a 3% grid step was profitable across every coin and period tested. Then live trading revealed the one thing the backtest underweighted: regime. Here's both halves, honestly.
I Built a Solana Sniper Bot in Python — Here's the Real P&L (Losses Included)
A fully automated pump.fun sniper bot in Python, funded with real SOL and run live. The architecture, the bug that killed 74% of buys, and the honest result — including why a 0.78-AUC model still didn't make money.
Build Log — June 10, 2026
Today's trading bots saw mixed results, with the Alpaca Crypto Grid bot finishing roughly flat after a day of active trading. The Kraken Grid + Swing bot, on the other hand, sat out entirely and also ended up roughly fla