One person and coding agents built a live trading system. The workflow that builds it grew with it.
I operate an automated trading system on Solana with my own funds. Coding agents built it through a delivery workflow that started as a nine-line file and grew, run by run and failure by failure, into a system of its own. Two case studies: the product, and the workflow.
Personal engineering project. It trades only my own funds: no client assets, no investment service, no advice.
Overview 0:50
The video could not load. Try Play again, or use the Open video link.
Read the transcript
I built SKYNET to bring trading intelligence and execution into one platform.
It tracks Telegram signals and on-chain wallet activity, with rankings that make their historical performance easier to compare.
Wallet profiles reveal trading results and behavior, while market intelligence adds context about changing conditions.
The strategy engine lets you replay historical signals with different exit rules and position sizes.
The trading terminal brings together Phantom wallet swaps, automated copy trading through my custom Rust engine, and live position monitoring.
Detailed transaction timelines show the journey from the source signal through execution to on-chain confirmation.
SKYNET connects the research, the trade, and the evidence behind it.
I own the product intent, the system boundaries and constraints, the acceptance criteria, the release decision and production verification. AI agents provide much of the implementation-level technical depth: they research, propose designs, write the code and review one another independently. I resolve their disagreements through evidence and decide what ships. I do not claim the unaided technical depth of a senior backend or Rust engineer.
Populations and datesShow less
A trade here is one position. 2,843 on-chain signatures stand behind the 1,300 positions, 1,685 buy legs and 1,158 sell legs, because a position can close in several legs. Engine time is the engine's own work per trade, from price quote to submitted transaction; 66 ms is the median across the 1,228 buys with stage timings, 90% of them under 200 ms. From signal to confirmation on chain the median is 1.3 s, most of it the blockchain's own confirmation (625 ms median) and the message's delivery from Telegram. Runs from 26 March to 1 September 2026, numbered 1 to 134. Measured 3 September 2026.
The system
A production system that turns public market signals into capped, auditable on-chain trades, with my own funds only: 66 ms of engine time per trade, about 1.3 seconds from signal to confirmation. Two servers, a custom Rust engine, and on-chain limits a trading key cannot bypass.
How I ship
One AI coordinates the job; specialist agents research, plan, build and test it; models from two other vendors challenge the work before anything risky ships. Ten phases, 62 steps, 34 gates, and a memory loop that turns failures into rules, across 112 documented runs.