About TradeBytes Ideas
A multi-agent AI research pipeline running 24/7 on local hardware. The goal isn't to publish a steady stream of picks — it's to wait patiently for asymmetric setups (think SanDisk's 1400% in a year) and only publish when something genuinely clears a high bar.
The pipeline
Five logical stages, two physical models, single-slot inference, file-based handoffs:
1. Hunter
MiniMax 2.7
Scans for catalysts: M&A rumors, insider clusters, unusual options activity, 13F changes, regulatory shifts. Surfaces events, not names.
2. Scout
MiniMax 2.7
Deep multi-hour investigation per symbol. Many tool calls — full 10-K and 10-Q reads, four most recent earnings transcripts, competitor filings, insider Form 4 history, options chain & IV, news, sector context.
3. Analyst
Qwen 3.6 35B
Reads the dossier, decides if there's an actual edge. Picks structure (long stock, cash-secured put, or covered call). Emits thesis, entry, exit, risk. Most often emits "skip — no thesis worth telling."
4. Devil's Advocate
MiniMax 2.7 (fresh ctx)
Adversarial review. Re-fetches sources to find evidence the Analyst missed or misread. Kills aggressively. Different model family from Analyst by design — diversity matters more than model size for breaking arguments.
5. Compliance & Publish
Qwen 3.6 35B
Disclosure check, factual claim audit, per-symbol cooldown, manual review for the first publishes. Then the static site rebuilds and Cloudflare Pages deploys.
Hardware
One Mac Studio M3 Ultra (256GB unified memory) running LM Studio. Both models pinned at 192K context. Single slot per model, no parallel sessions — jobs queue up and the box runs continuously. Editorial pipelines for the news site (tradebytes.net) share the same hardware.
What we're optimizing for
- Patience. Default output is silence. The system is allowed to research a symbol for hours and return "no thesis." That's a feature.
- Depth. A dossier with 40 tool calls beats a dossier with 4. We're hunting for what nobody else read.
- Transparency. Every published idea names the models and shows the Devil's Advocate verdict. Every close is logged on the track record page.
- Discipline. Position size is suggested, time horizon is fixed, stop conditions are explicit. No moving the goalposts after the fact.
What we're not
- Not a registered investment advisor.
- Not personalized advice — we don't know your situation, goals, or risk tolerance.
- Not affiliated with any broker, custodian, issuer, or third party. No paid promotions, ever.
- Not infallible — an AI thesis can be wrong, even after deep research and adversarial review.
Read the full disclosures before relying on anything published here.