Soubhagya Swain@soubhagya_swain881
2
Events attended
1
Submissions made
India
1 year of experience
About me
I enjoy working close to the metal, thinking through how things behave under the hood, and building software that helps people. Most of my interest sits at the intersection of systems thinking, data, backend logic, and developer tooling. If it is efficient, well structured, and a little obsessive about detail, I will probably like it.
🤓 Latest Submissions
MergeCalibr-Jev
MergeCalibr-Jev is an AI-powered pull request triage system that helps development teams cut through the noise created by high-volume, AI-assisted code contributions. As AI coding tools accelerate how fast code gets written, human reviewers are left struggling to keep pace — every PR looks equally polished, making it hard to tell which changes are genuinely safe to merge and which need careful scrutiny. The system works in two distinct stages. First, a large language model reads each pull request's diff and produces a clear, structured summary of what actually changed — which files were touched, what the functional impact is, and which parts of the codebase are affected. This step turns a raw code diff into something a decision system can reason about. That summary is then passed to a calibrated decision model, which evaluates it against a fixed set of typed questions — how risky is this change, which subsystem does it touch, and does it show signs of a security flaw — and returns fast, mathematically calibrated probabilities rather than free-form text. Because these confidence scores are calibrated, the system can safely make automated decisions instead of relying on a model's subjective-sounding opinion. Based on those scores, MergeCalibr-Jev routes each PR down one of several paths: low-risk, high-confidence changes are fast-tracked straight to normal review; medium-risk PRs are sent to a human reviewer along with the generated summary and flagged concerns, so they aren't starting from a blank diff; and high-risk or security-flagged PRs are automatically escalated to an AI coding agent, which audits the flagged code against defined security rules and applies a fix before a human reviewer ever sees it. The result is a workflow where human attention is spent only where it's actually needed — reducing review fatigue, shortening the time code spends waiting in review queues, and catching risky changes before they reach production, without slowing down the safe, everyday work that makes up most of a team's pull requests.
27 Sep 2026
