Fully automated Python trading engines wired to live news feeds, trained on real trading-class material, and journalled to a web dashboard you can open from anywhere. The bot trades. The journal proves it.
The trading core is pure Python — connects to your broker API, runs the strategy loop, executes orders, handles retries and reconnects. Runs 24/7 in a Docker container with auto-restart, no babysitting required.
Pulls real-time market headlines and macro events from sources like Polygon, Alpha Vantage, NewsAPI, and Benzinga. Sentiment is scored and fused with price action so the bot trades on what's actually happening, not yesterday's chart.
Your course notes, mentor playbooks, and the psychology behind disciplined trading are ingested as the bot's knowledge base. The model learns rules it can quote — entries, exits, risk caps, position sizing — not just patterns it guessed at.
Every move — entries, exits, the news the bot saw, the reasoning it gave — is written to a clean web dashboard. Open it on any device. Search, filter, replay. No CSVs, no terminal logs. Your performance history, readable.
A nightly review loop replays the day's trades, scores what worked and what didn't, and adjusts confidence weights for setups, time-of-day, and news categories. The bot gets sharper at reading your market, not a generic one.
Two brains, your call. Hosted LLM (Claude / GPT) with a RAG layer that reasons in natural language and trades like a disciplined human, or a classical ML model (gradient boosted trees, LSTM) for pure-signal latency-sensitive setups. Either way, the Python execute engine has final say.
Price ticks from the broker, headlines from news APIs, macro calendar events. Normalised into a unified stream with timestamps, source, and confidence.
polygon.io · alpha_vantage
Retrieves relevant rules from your trading class & mindset notes. Reads the news, scores the setup, proposes an action with a written rationale and confidence.
claude · gpt · vector_db
Validates the trade against hard rules — position size, risk caps, market hours, instrument whitelist. Rejects or trims anything the LLM hallucinated.
execute_trade.py
Order writes to the broker. Every detail — inputs, reasoning, fill, P&L — writes to Postgres. The web journal updates live. The nightly loop reviews.
postgres · web_journal
LLMs hallucinate. They can invent ticker symbols, miscount lots, or talk themselves into oversized positions. That's fine — because the LLM never touches your broker.
Every action the model proposes is parsed, schema-validated, and run through a deterministic safety engine in Python before a single order goes out. If the LLM says "buy 4,000 shares of TSLAA", the engine sees: invalid ticker, position exceeds risk cap, market closed — and the trade dies on the floor with a logged rejection. The bot never hesitates, never freezes, never wires up a hallucination to real capital.
The bot writes a structured entry for every action it considers — not just fills. Skipped setups, rejected hallucinations, take-profits, partial closes. You see what it saw, why it acted, and what it learned. Filterable, searchable, exportable.
Headline at 14:28 confirmed enterprise GPU contract. Volume spike on 5m matched the rule from week-6 lesson notes ("breakout on news + above 20MA"). Position sized to 0.6% account risk; stop $866.20.
CPI print 0.2% above consensus. RAG retrieved the macro-surprise playbook; brain proposed short with high confidence. Safety engine confirmed risk fit, executed. Trailed stop on 1m close above VWAP.
LLM proposed long 80 shares of "TSLAA" — symbol not in whitelist (likely hallucinated). Trade rejected before reaching the broker. Logged for model review. No capital exposure.
Pre-market gap with no negative news matched the "gap-and-go" framework from class. Entry confirmed by 9:45 volume bar. Held for first profit target.
Setup looked acceptable but model confidence under the threshold. Discipline rule: when in doubt, sit out. Logged so the learning loop can review whether the threshold needs adjustment.
A hosted Claude or GPT model with a retrieval layer over your trading class, mentor notes, and historical journal. It reads the news, recalls the rule, writes the rationale, proposes the trade.
Classical machine learning — gradient boosted trees, LSTM, or a custom ensemble — trained on your historical signals and labels. No language, no narration. Pattern in, probability out.
// You can also run both — LLM for daily macro, ML for intraday execution. I've shipped that hybrid before.
I sit with you, your class notes, and your existing playbook. Every rule becomes a testable spec.
Historical price + news data assembled. Strategy back-tested across multiple regimes before any live capital.
Bot runs live but trades a simulated account for 2–4 weeks. Journal proves it follows the rules under real conditions.
Capital allocated in stages. Safety caps tight at first, widened as the journal shows consistent behaviour.
Nightly review loop tunes confidence weights. Monthly model retraining. New class material rolls in as RAG context.
Tell us your edge — the setups, the rules, the discipline you trust. I'll engineer the engine that runs it 24/7 without flinching.
Start the Build