Developer's AI PR Reviewer Hits 97.7% Accuracy After Fixing the Wrong Bug for Weeks

·1 views

A developer building an AI-powered pull request review plugin discovered that the baseline model, Claude Opus with no added guidance, already catches around 65–70% of common code bugs on its own. The plugin's real advantage came not from finding more bugs but from suppressing false positives and accurately classifying risk levels across different code change types. During evaluation across 43 test scenarios spanning four custom repositories, the plugin reached a 29.5 percentage point accuracy gap over the baseline after harder test cases were introduced. One of the key lessons was that reweighting the evaluation metric for risk classification from 5 to 10 points widened the gap by 9 percentage points without any code changes. The developer also spent several iterations rewriting the wrong component before realizing the actual fix was a single-line change in an upstream skill.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Five Chrome Extensions That Export ChatGPT Chats to Word or PDF Files

Several Chrome extensions now allow users to export ChatGPT conversations directly into Word or PDF files, solving the common problem of lost formatting when copying text manually. ChatGPT lacks a built-in export feature, making dedicated browser tools useful for writers, students, and office workers who rely on it daily. Key factors to consider before choosing an extension include supported output formats, export scope, formatting fidelity, and workflow compatibility. One highlighted option, ChatGPT Exporter, supports both Word and PDF export, preserves tables, code blocks, and math equations, and also integrates with Google Docs and Notion. The guide reviews five such extensions in total, each suited to different working styles and export needs.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

APSentra Models Organizations as Live Graphs to Fix Enterprise Procurement Workflows

Software firm APSentra has published details of an architectural approach that treats company structure as a queryable, versioned graph rather than a static configuration to drive procurement approvals. Traditional workflow engines fail at enterprise scale because they hardcode approval sequences that quickly become outdated as organizations change, the team argues. In their system, employees, roles, cost centers, budget pools, and reporting relationships are stored as nodes and edges in a graph database, with every structural change recorded via timestamped edges rather than overwriting existing records. This design allows approval chains to be computed dynamically at runtime by traversing the graph, rather than relying on deployment-time workflow definitions. The versioned graph also enables precise audit queries, such as determining exactly who held approval authority on the specific date a procurement request was submitted.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer's 3-Week AI API Price Audit Finds Cheap Open-Source Models Rival Costly Ones

A developer building on large language model APIs for over three years conducted a three-week audit of AI API pricing after noticing token costs exceeded server costs on their AWS bill. Comparing dozens of models and normalizing prices using data from Global API's pricing endpoint verified in May 2026, they found the cheapest viable models are largely MIT and Apache 2.0 licensed open-source models, primarily from Chinese model families. The analysis revealed a pricing spread of up to 40 times between models with only marginal benchmark differences, such as 88% versus 92% on MMLU. The author organized models into cost tiers ranging from under $0.10 to over $0.80 per million output tokens, arguing that open-source alternatives have quietly caught up with — and in some cases surpassed — proprietary incumbents. The findings challenge the widespread developer assumption that high-quality AI necessarily means high-cost AI.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developers Explore AI as Active Meeting Participant, Not Just a Note-Taker

A developer writing on DEV Community argues that current AI meeting tools focus too narrowly on recording and summarizing conversations rather than reducing the follow-up work meetings generate. The author proposes a shift toward AI agents that join meetings in real time as active participants rather than passive recorders. They have been experimenting with this concept using a tool called AgentCall.dev, which they say feels more like collaboration than traditional meeting assistance. The piece suggests the future of meeting technology lies in AI that participates in discussions rather than merely documenting them. The author invites other developers to share their perspectives on this evolving approach.

0 comments Read more at DEV Community