Rodrigo Fernandes
One of the systems below has traded real money on Solana. AI writes a lot of the code. Judging it is my job.
A Helius webhook listener, a position monitor, and a Telegram interface ran around the clock, each able to wake a Claude agent that acted through 25 tools. It wrote its own lessons, retuned its own runtime config, and once closed a position at 5 a.m. with nobody awake. The guardrails live in tool code, not the prompt: hard position caps, a daily-loss circuit breaker, no transfer tool.
An on-chain protocol demanded time-sensitive actions from its users, so I built an agent to take them over. Users fund a dedicated Privy server wallet and set hard limits, then a worker loop sizes every action in deterministic BigInt math, simulates each step against the live chain, and won't call anything done until on-chain state confirms it. The LLM explains strategies; it never executes.
Four specialist agents audit a research paper for math, logic, methodology, and reference errors. An orchestrator classifies the paper and decides which agents read which sections through per-agent RAG retrieval, every finding lands with section and page provenance, and an eval harness injects ten known errors into real papers to measure detection recall.
Designed and shipped a full EVM protocol to mainnet: eight Solidity contracts of custom token economics, priced by a Uniswap V3 TWAP oracle I ported to Solidity 0.8 after an audit finding. Around the contracts sit 233 Foundry tests, a 47-finding internal audit tracked to zero open highs, a five-duty keeper, and a market-making bot swarm.
A pricing engine for a collectibles marketplace that refuses to guess. Comps from eleven sources pass a seven-field identity match and cross-feed deduplication before a recency-weighted median sets fair value, and under six verified comps it returns insufficient-data instead of a verdict. A weekly backtest replays its own decision log and reports the realized hit rate.
A capsule endoscopy exam produces over 12,000 images, and reading one takes a gastroenterologist 30 to 40 minutes. My thesis attacked the annotation bottleneck: a ResNet50 Siamese tracker expanded expert labels across neighboring frames to 72,000 annotated images, and when tracking under-performed I pivoted to YOLOv5x detection across six lesion classes, reporting the failures as carefully as the wins.
Start from what's actually running: the repo, the logs, the data model, the external services. Name the exact failure mode before changing anything.
The agent gets a fence before it gets a task: only the files and behavior the change touches. A narrow brief is the cheapest guardrail.
AI drafts the patch. Architecture, naming, and product logic stay under my review, and nothing ships that I can't explain.
Run the full flow end to end, from API to dashboard to webhook, then read the deployment logs before calling it done.
If this is how you want software built, let's talk.