EvidenceGate Memory Agent

EvidenceGate uses Qwen Cloud and SQLite memory to review AI-generated analyst reports, flag missing evidence, and return a clear REVISE verdict before trust.

Comment

Inspiration

AI-generated analyst reports can sound confident even when important evidence is missing. In trading research, this is risky because a report may mention Entry, SL, or TP without showing enough support from higher-timeframe context, invalidation, volume, Fibonacci, or remembered review rules.

EvidenceGate was built to make memory enforceable. Instead of letting an AI report pass because it sounds fluent, the system checks whether the report contains the evidence required by persistent review memory.

What it does

EvidenceGate is a memory-backed evidence review layer for AI-generated analyst reports.

It uses Qwen Cloud to extract structured evidence from a report and generate embeddings. It then retrieves persistent review rules from SQLite memory and compares the report against those rules.

If required evidence is missing, EvidenceGate returns a clear REVISE verdict with the missing evidence listed.

The demo domain is trading research, but EvidenceGate is not a trading bot. It does not issue live trade signals or approve trades. It reviews whether an AI-generated report is supported enough to be trusted.

How we built it

We built EvidenceGate with a static HTML/CSS/JavaScript frontend, a Python HTTP API backend, SQLite for persistent memory, and Qwen Cloud for evidence extraction and embeddings.

The backend sends analyst report text to Qwen Cloud, stores and retrieves review memory from SQLite, then applies local evidence-gate logic to produce a structured audit result.

Challenges we ran into

The main challenge was avoiding overclaiming. We kept the project focused on a working MVP: evidence extraction, memory retrieval, and a clear review verdict.

Another challenge was making the demo understandable without presenting the system as a trading engine. The final framing is intentionally narrow: EvidenceGate reviews AI-generated reports, not markets.

Accomplishments that we're proud of

We built a working memory-backed review flow using Qwen Cloud and SQLite.

The demo shows an AI-generated analyst report being checked against remembered evidence requirements and flagged with a REVISE verdict when support is missing.

We are also proud that the project stays honest about what it does. It does not claim to predict markets or approve trades. It focuses on evidence quality.

What we learned

Memory becomes much more useful when it is connected to enforceable checks.

For AI-generated analysis, retrieved memory should not only provide context. It should also help decide whether an output is sufficiently supported.

What's next for EvidenceGate Memory Agent

Next steps include adding more review rule sets, supporting additional report domains beyond trading research, improving the UI for reviewer workflows, and deploying the backend as a hosted cloud service.

Built With

Share this project:

Updates