The Sidekick code review team at HubSpot utilizes OpenAI's Decisions API to streamline triage by asking specific questions about pull request properties for quicker evaluations.
added by @zklapow
At HubSpot, our Sidekick code review team started using OpenAI's Decisions API for triage: instead of asking a full agent "does PR get a security review?", we instead ask questions about properties of a given PR, using the decisions API to get fast answers on what to evaluate.
View similar
Ranked from stored criteria vectors. No live classification on this page.
TypeSafe AI's Jev model has been integrated into a system layer, providing clean, typed decisions from unstructured state in under 500ms. This integration enhances backend processes by scoring and filtering context efficiently.
Developed a qbash Task to match company names from call transcripts to a database, utilizing phonetic encoding to refine options. Achieved an estimated model cost reduction of approximately 84% with a minor accuracy tradeoff.
This approach involves using Codex to create APIs for web applications, which are then connected to ChatGPT via custom GPT Actions. With GitHub access to the project code, ChatGPT can provide insights about the app's architecture and behavior.
The Jev model was benchmarked against an ensemble of models for code reviews, achieving zero false positives, a review speed increase of ~50x, and a cost reduction of ~100x. It demonstrated a 75% bug recall rate, highlighting its efficiency compared to traditional multi-turn agent workflows.
The Spectra system utilizes Deepgram Flux for speech-to-text functionality, providing end-of-turn reports with probabilities. This setup ensures high accuracy and quick responsiveness for agent interactions.
This process involves handpicked RSS feeds and an automated curator that evaluates date-bound feeds. A custom set of heuristics is used to score and rank posts, with top selections identified through a classifier.