The scanner is a research system, not a prediction machine
When I started building Aventra Market Radar, I did not want to create a black-box signal generator. My goal was to reduce the time required to review thousands of charts while preserving the trader's responsibility to understand context, invalidation, and risk.
That decision shaped the architecture. Pattern detection remains rule-based and explainable. AI strengthens the surrounding research workflow: it helps organize candidates, connect market context, explain why a structure was surfaced, and make a large result set easier to investigate. It does not replace the structural rules or turn a candidate into a guaranteed trade.
Start with a canonical market universe
A scanner is only as trustworthy as the symbol universe behind it. Provider symbols, exchange symbols, and the symbols shown to users are not always identical. A suffix may be required by one data source but invalid in a charting link; the same company may also appear through different market conventions.
I therefore treat the application market universe as a separate product domain. Canonical symbols, countries, exchanges, memberships, benchmarks, and provider mappings are resolved before scanning begins. This prevents a data provider from silently redefining what the product considers active or scannable.
- Keep canonical and provider-specific symbols separate.
- Resolve benchmarks by market and timeframe.
- Do not let partial provider responses add or remove product symbols implicitly.
Separate detection, quality, and presentation
A visual resemblance is not enough to call something a useful pattern candidate. The detection layer asks whether the required pivots, boundaries, ratios, trend conditions, or setup events exist. The quality layer evaluates clarity, recency, volume behavior, relative strength, trigger proximity, and invalidation. The presentation layer then draws only the information a reviewer needs.
Keeping these responsibilities separate makes the system easier to test. It also means a chart can explain the same geometry used by the scanner instead of relying on an unrelated drawing routine. When a detector changes, I can evaluate its effect on candidate counts, chart annotations, and historical performance without rewriting the entire product.
Make expensive work incremental and failure-tolerant
Market scanners become operationally expensive when every refresh recomputes every historical bar, every feature, and every symbol. Aventra uses bounded refresh windows, overlap for late corrections, latest-feature snapshots, bulk writes, and incremental updates that focus on symbols receiving new bars.
Partial failure is equally important. A missing or temporarily unavailable symbol should not discard successfully ingested data for the rest of an exchange. Pattern Search and Ready Setups can run independently, and scanner results are persisted in batches so users are not left with an empty product while a long refresh is still running.
Why human review remains central
Two candidates with similar scores can have very different market context. Earnings, corporate actions, liquidity, nearby support and resistance, and the quality of the broader trend still matter. That is why Market Radar includes table review, multi-chart review, detailed chart drawings, watchlists, backtesting, and performance observations rather than one oversized buy or sell label.
For me, the practical role of AI in technical research is to make disciplined analysis faster and clearer. The product should help a person ask better questions, compare more evidence, and document invalidation—not hide uncertainty behind a confident prediction.