Skip to content
MDM
2026

Active-learning ranking app

Learns preferences from head-to-head choices with a Bradley-Terry model, and always asks the comparison it is least certain about. Built on the database from my agent system.

Kind
AI system
Context
Personal project
Result
A ranking I trust that decides where I spend application time, and sharpens with every choice.
Built with
Python, SQLite, Bradley-Terry, MM algorithm, HTML/JS

The problem

With hundreds of interesting roles, scoring each one from 1 to 10 gets inconsistent. Comparing two roles at a time is much easier, and more honest.

What I built

A web app on top of the job database from my agent system, reachable only from my own devices.

  • New postings found by the agents appear in the app with a summary, requirements, deadline and fit notes.
  • Ranking by choice. I pick the better of two roles. A Bradley-Terry model, fitted with the MM algorithm, turns those choices into Elo-style scores. The app then picks the next pair where the ranking is least certain, so every comparison teaches it as much as possible.
  • Useful from day one. New roles start from a prior, the agent's 1–10 fit estimate, until I compare them. The ranking gets sharper with every choice.
  • Around it: rankings per region from one global model, statuses and notes, undo, sync to my notes, and an Applications tab where I request a tailored application pack and give feedback on each item.

The model in one line

Bradley-Terry gives every role a strength s. The chance that role i beats role j is si / (si + sj). The MM algorithm finds the strengths that best explain all the choices so far.

Compare screen: two roles stacked, each with why it fits, and Undo, Equal and Skip buttons between them
Compare: pick the better of two roles.
Ranking screen: a numbered list of roles with a score and an uncertainty band for each
Ranking: score and uncertainty per role.
Jobs screen: today's new roles with Shortlist, Apply and Not for me buttons
Jobs: new postings from the agents.
Applications screen: a list of applications with statuses such as Draft, Interview and Rejected
Apply: every application and its status.

All screenshots use demo data; the real roles and companies stay private.

Result

A ranking I trust. It decides where I spend my application time, and it improves with every choice I make.

The model can be checked too. On a demo with 30 made-up roles and 150 choices simulated from known hidden strengths, the fitted order reaches a Spearman correlation of 0.81 with the hidden order. I also split off a standalone version that ranks any list from a CSV file.