End-to-end automation: three systems that watch markets and send alerts on their own
Three Python systems that automate market tracking: written rules, email alerts and a review after every close. They run on paper or only suggest: the final decision is mine.
What it automates
Following several markets by hand, with rules that change from day to day, doesn't scale. The systems apply written rules, send email alerts, review themselves after every close and don't execute a real order without a decision from me.
The three systems
| System | What it does |
|---|---|
| Paper bot | Weekly screening of the universe, evaluation across several timeframes, simulated trades and a daily email. |
| Accumulation core | For five large assets: confluence zones, buy tranches, confirmed breakout and monthly allocation. It suggests by email. |
| Vigía | Alerts on a real portfolio by reading the API in read-only mode: stops, conviction, candidates and status. |
How it works
- 1 · Data
Binance, Yahoo, CoinGecko. eToro: read-only.
- 2 · Universe
Weekly screening and re-scan, with atomic writes.
- 3 · Evaluation
Filters in fixed order: validity, cycle, grade, signal, room, R/R.
- 4 · Alert
Daily email via SMTP: how each filter fared and which came closest.
- 5 · Human decision
No real order without a person.
- Post-close reviewer
Evaluates each trade at +2 and +7 candles (in ATR units). Proposes thresholds: never edits the configuration.
- Three-layer audit
5 code auditors, 3 adversarial verifiers and 3 experts. No finding was refuted.
- Daily summary with source
Explains what moved and why, with source and date. If the AI fails, the email goes out anyway, without AI.
The lines connect what feeds to what reviews. The only step that executes anything is step 5, and a person does it.
How it works unsupervised
- Eight scheduled jobs synchronized with the daily candle; the monthly ones run on the 1st.
- State with atomic writes and file locks, and retries that don't duplicate. A network error is not mistaken for “no signal”: on the third failure in a row, a warning email arrives.
- A three-layer code audit: code auditors, verifiers who tried to refute each finding, and experts. Everything that surfaced was fixed in groups, each with its own commit.
- A reviewer that proposes and doesn't edit. It evaluates each trade at +2 and +7 candles and requires a minimum sample per group before suggesting a threshold.
- A summary by email every day, with what happened and what to look at.