01
Market data
Streams order books and trades from Coinbase, Kraken, OKX, Hyperliquid and Kraken Futures over WebSockets, rebuilds the books, and stores everything in a Parquet data lake partitioned by venue, symbol and day.
Vigilo · Latin for “I keep watch” · a personal project by Dawit Weldemichael
A real-time trading research platform for crypto and prediction markets.
Vigilo records order books and trades from five crypto exchanges, tests trading strategies against live prices on Kalshi, a CFTC-regulated prediction market, and runs a strategy with real money under hard risk limits. One engineer designed, built and operates it, on an AWS server, around the clock.
The live system is shared by personal link.
If you received a link or QR code, open it to see the platform running: live data, strategy results and the operator dashboard, read-only. To request access, get in touch on LinkedIn.
What it does
Four systems, one platform
02
Strategy research
Paper-trades strategies against real Kalshi prices with conservative, queue-aware fill simulation. Results are judged on return per dollar and statistical significance, not win rate.
03
Live execution
Copies a research strategy onto Kalshi with real money: post-only orders, hard per-bet, total and daily-loss limits set, duplicate-safe order handling, and reconciliation against the exchange's records.
04
Operations
Independent services that restart on failure, a multi-user strategy lab with role-based access, and an operator dashboard for monitoring everything in real time.
AI
AI built in, with guardrails
Vigilo uses AI in two ways: agents that watch and analyse the system, and an assistant for every strategy researcher. Both work from real data through narrow, read-only tools.
Sentinel
Watchdog
A scheduled agent that checks every service, data feed and trading bot several times a day, and raises the alarm with a suggested fix.
Quant
Analyst
Reads the day's paper trades and arbitrage logs and writes a daily briefing on what worked, what didn't and why.
Scout
Edge hunter
Every week it looks for a new trading edge by itself: it forms hypotheses, runs quick backtests, and submits its best idea to a test on hidden recent data that code, not the AI, grades. It also answers questions on request.
Assistant
Strategy Lab helper
One per user. It inspects their strategies, runs backtests and drafts changes through a tool-use loop, sees only their own data, and stays within per-user cost limits.
The agents are headless AI with a read-only toolset: they can't edit files, restart services or trade, and nothing they report is acted on automatically. This is deliberate: an AI model can misread data, or be misled by text it reads, such as a market title or a log line written to trick it. So the agents can advise but never act, and every change goes through a person.
How it works
From exchange WebSocket to trading decision
The services share data through files rather than network calls. The collector writes, and everything else reads, so a failure in one strategy can't stop data collection or real-money risk checks.
- SourcesFive crypto exchanges over WebSockets, plus Kalshi for prediction markets.
- CollectorOrder-book rebuild, resync on sequence gaps, clock-drift checks, batched writes.
- Data lakeParquet files queried with DuckDB and compacted nightly; SQLite for engine state.
- EnginesPaper traders, a strategy lab, a market maker, an arbitrage scanner, research studies and the live trader.
- OutputAn operator dashboard and alerts for fills, results and outages.
Python · asyncioFastAPIDuckDBParquetSQLiteWebSocketspytest · mypy · ruffClaude API · tool useClaude Code agentsEChartssystemdCaddy · TLSAWS EC2
Interested in seeing it run?
I'm happy to share a personal access link or walk through the architecture, the research and the risk controls.