Vaibhav “VB” Srivastav works on developer experience at OpenAI, helping developers make Codex effective, configurable, and safe for everyday software engineering. Previously head of developer experience and community at Hugging Face, he built open-source tools and research spanning speech recognition, inference optimization, and machine-learning evaluation.
From open speech infrastructure to software agents
During approximately three years at Hugging Face, Srivastav worked across audio, on-device machine learning, developer education, model integrations, inference providers, and community programs. His Hugging Face portfolio includes models, datasets, interactive applications, and developer experiments.
He created Insanely Fast Whisper, an optimized transcription tool that makes Whisper-based speech recognition more practical to deploy. As first author of research introducing the Open ASR Leaderboard, he helped establish reproducible comparisons of open and proprietary transcription systems across languages, recording lengths, accuracy, and processing speed.
His contributions also include ESPnet-EZ, a Python-oriented interface for speech-model fine-tuning and inference, and technical writing on self-speculative decoding, which accelerates generation by having earlier model layers draft tokens that deeper layers verify. He additionally coauthored the Transformers v5 announcement, covering simpler model definitions, interoperability, and production deployment.
- Evaluate speech recognition under real deployment constraints. Transcription quality means little without comparable measurements of speed, multilingual performance, and behavior on long recordings.
- Treat faster inference as developer access. Efficient transcription, practical fine-tuning, and lower-latency generation make capable models usable in applications with limited compute or demanding response times.
- Make code review repository-aware. Srivastav advocates reviewing changes against the surrounding codebase, identifying downstream consequences outside the immediate diff, and prioritizing findings by severity.
- Give custom subagents bounded authority. Specialized reviewers, researchers, and implementation agents can divide independent tasks while receiving different models, reasoning budgets, tools, and sandbox permissions. Security and review agents should remain read-only; agents responsible for writing need narrowly appropriate access.
- Keep privileged actions accountable. Approval mechanisms, event-driven hooks, and restricted execution environments can reduce repetitive supervision without granting software agents unrestricted control.
At AI Engineer Europe 2026, Srivastav explored these practices alongside Katia Gil Guzman in a Codex and subagents workshop, personally demonstrating repository-aware review, configurable agent personas, sandbox controls, and approval workflows. His public update on Codex session-efficiency improvements extends that hands-on focus to operational issues affecting working developers.
