Skip to content
Align your AI

Research

Continuum: Automated Construction and Retrieval of Software Decision Knowledge Graphs from Developer-AI Conversations

A published system paper on capturing the decisions made in chats between developers and AI coding assistants, and serving them back to people and coding agents.

Venue
ASE 2026 Industry Showcase
Authors
Mohammad Ali Shehral, Karthik Ravi, Nikhil Trivedi and Akram Bayat

Talk

Continuum: Decision Knowledge Graphs from Developer-AI Conversations (ASE 2026)

ASE Industry Showcase

Playing from YouTube; their cookies and requests start when you press play.

Abstract

Developers now make many technical decisions in chats with AI coding assistants, and the reasons are lost when a chat ends1. Continuum captures each decision from chat logs, with its trigger, options, choice and rationale. It resolves the many names one technology goes by and stores the decisions in a knowledge graph.

People search it, and coding agents query and update it through the Model Context Protocol. The paper's stated contribution is this end-to-end integration, not a new algorithm1.

What this shows

  • On a synthetic corpus of 200 generated conversations, the name resolver reached 97.2% accuracy on 981 test mentions, against 74.6% for exact matching1.
  • The pipeline ran end to end, including on the chat history of one real project1.
  • What counts as a decision is hard to agree on. On the 10 items that two authors both judged, agreement was close to chance, a Cohen's kappa of 0.07. The paper reads this as a finding about the task, not only the tool1.

What this does not show

The paper lists these limits itself1.

  • Performance on real team conversations. The main corpus was generated by AI, and the real-log case study covers a single project1.
  • Fully independent ground truth. For 150 of the 200 conversations, the answer key partly came from the system's own resolver, which may inflate accuracy1.
  • Retrieval on natural questions. The 40 retrieval queries were built from the graph, which biases recall upward1.
  • That it beats other tools, or that teams decide better with it. No user study or head-to-head comparison was run1.
  • Privacy at team scale. The case study's scrubbing relied on manual review that does not scale1.

Sources

  1. Shehral, Ravi, Trivedi and Bayat, “Continuum: Automated Construction and Retrieval of Software Decision Knowledge Graphs from Developer-AI Conversations”, ASE 2026 Industry Showcase. https://doi.org/10.1145/3832783.3834535. Accessed 2026-10-13. ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11