Self-directed · AI-assisted · Local LLM · 2026

Deal
Analyzer

A local model watches the deal feeds I can't, and weighs each one against the calls I've already made.

The Deal Analyzer review queue, shown under its internal build name: each deal displays a good-or-bad verdict, a confidence percentage, how long the model took, its written reasoning, and buttons to confirm or correct the call.
Fig. 1: Verdict, confidence and reasoning per deal. The interface still carries the internal build name.
Alert
~20 sec
Escalates
Below 85% confidence
Runs
Locally
Cost
$0 to run

Problem

Good deals surface and disappear inside minutes, and I can't sit and refresh a feed all day. A fixed keyword rule was no help; whether a deal is good depends on resale price after fees.

Evaluation

Two tiers. A fast local pass returns a straight good-or-bad verdict, which handles most deals. When confidence falls below about 85%, the model escalates to slower, more thorough reasoning on market value, resale margin and fees. The fast pass minimizes delay for routine calls, while the deeper pass spends additional time only on uncertain cases. An alert typically lands around 20 seconds after pickup, while the deal still exists.

Using past decisions

When I approve or reject a recommendation, that decision is stored as a historical example. When evaluating a future product, the system retrieves relevant past calls, such as recent approvals or denials involving the same CPU family, and supplies them as context in the evaluation prompt. Roughly ten examples get pulled in, depending on the product category and how well the matching lands.

Delivery

Verdicts stream into the UI as they land and anything marked good pings me instantly. Processing is designed to keep pace with the source's normal posting rate. It runs on my own hardware, so evaluating everything costs nothing per deal.

Stack

AI-assisted development Ollama · local LLM FastAPI · SSE Playwright SQLite · SQLAlchemy React · TS · Vite Discord webhooks