Mike Beaumier

Designed and implemented company-wide AI code review tooling2026

• Designed, built, evaluated, and launched an AI code reviewer in GitHub that took first-round review off engineers and cut time to approval 41-56% on size-matched changes by ending drawn-out multi-round review threads
• Cut the AI code reviewer's cost to about a tenth of an engineer's review by redesigning it around one higher-quality agent
• Designed and built a benchmark of human-reviewed pull requests, stratified by coding style, size, and domain, showing an AI code reviewer running without developer oversight matched human findings at 89.1% precision
• Ran the head-to-head evaluation of AI code-review approaches that set the company's choice, picking a reviewer that caught nearly all known issues at a fraction of the cost

Architected data provenance metadata registry for all data artifacts2026

• Set the company's data-provenance direction by designing a metadata registry for every dataset and model, writing its build-versus-adopt analysis, and driving its adoption into the shared data platform

Transformed and hardened vulnerability detection workflows2025-2026

• Matched a deep-learning vulnerability detector for compiled code at 97.6% pairwise accuracy with gradient-boosted models on domain-knowledge features
• Designed and built an expert-in-the-loop annotation platform pairing structured LLM analysis (Pydantic AI) with expert review, which produced the company's gold-standard evaluation set for vulnerability detection in compiled code
• Showed LoRA matching full fine-tuning of a 110M-parameter transformer for vulnerability detection in compiled code while training 1.07% of its parameters
• Shaped the research direction toward agentic approaches with evidence from expert review, gradient-boosted baselines, and a LoRA study

Established internal AI engineering platform for all developer workflows2025-2026

• Designed, built, evaluated, and launched the company's AI coding agent platform, doubling engineering coding output
• Designed and built the agent workflow that takes a ticket to a merged pull request, delegating each stage to shared skills
• Designed and built a single source format for the company's agent skills, transpiled per editor, so each skill installed correctly in Claude Code, Cursor, and OpenCode without forking the library
• Authored the RFC for measuring whether coding agents follow the company's written instructions, separating whether an agent consulted an instruction from whether its output conformed

Redesigned and led engineering efforts to standardize all ML training operations as directed acyclic graphs2025

• Designed and built the research team's Dagster pipeline platform through production Kubernetes, cutting environment setup from days to minutes and letting researchers launch GPU training jobs themselves
• Kept the research team's self-service pipeline platform architecturally coherent as adoption grew, writing its guide and design doc and guiding teammates through review as they built 3 research pipelines on it

Multi-agent orchestration and agent memory

• Designed and built a multi-agent orchestrator that ran parallel coding agents (Claude Code, Codex, Gemini CLI) to plan, write, review, and validate code autonomously over multi-hour runs, before such tools were standard
• Separated real defects from style preferences in code from parallel coding agents by having a second vendor's model (Claude or OpenAI) review each agent's work and arbitrating disagreements before merge
• Built long-term memory for coding agents, a Claude Code plugin with session hooks and durable journals, so what an agent learned carried across sessions and context compaction instead of being rediscovered
• Built a service that condensed each coding-agent session (Claude Code, OpenCode) into a searchable, secret-redacted briefing with capped spend, so a person or another agent could resume after a crash or handoff

Gemini review agent for ad-policy reviewers2024-2025

A Chrome extension built on an agent architecture: it processed each webpage and surfaced annotations that pointed human content reviewers to the parts most likely to violate ad policy.

• Conceived, architected, and led the launch of a Gemini-powered LLM agent that flagged ad-policy violations for Google's content reviewers, used in 40,000+ reviews a day and raising throughput 30% in A/B tests
• Built and led the agent's cross-functional team of engineers, analysts, and product managers, and designed and built the A/B testing framework whose randomized controlled trials gated every launch

Quality measurement for LLM content classifiers2024-2025

• Guided the design and validated the launch of Google's LLM auto-labeling classifiers, trusted only where precision's 95% lower bound cleared 80%, scaling review 10x across billions of web pages
• Set Google Ads Safety's org-wide launch QA for LLM content classifiers, designing and building the sampling-and-review check for classifiers set to launch unchecked after tuning on dozens of examples
• Built precision dashboards for Google's production LLM content classifiers from a continuous stream of human-reviewed samples, using bootstrapped confidence intervals to flag underperformers for prompt retuning or retraining

Review efficiency for publisher safety2024

• Let Google's publisher-safety reviewers cover 30% more content at higher precision by leading a working group that cut reviews of pages that never served ads and set a cost-per-impact metric
• Raised enforcement up to 200x on some policies with high-precision rules running at over 80% precision

Automatic Classification of Suspicious Web Content2023

• Designed and built AutoML classifiers on SQL pipelines joining web-content signals with past human review outcomes, cutting manual review up to 80% across hundreds of millions of pages, saving millions a year

Ad Query Forecasting to Prioritize Content Review for New URLs2022-2023

• Led a traffic-forecasting system end to end that flagged high-traffic new URLs within 90 minutes, cutting their harmful ad impressions 60% in the first 24 hours, in a pipeline streaming hundreds of millions of URLs weekly
• Identified the 90-minute window after a high-traffic URL first appears in which most of its harmful ad impressions can be prevented, from time-series analysis of Google's ad-query logs covering billions of URLs

Machine learning fairness for content classification2022

• Led Google Ads Safety's ML fairness working group and set the org-wide bias metric for its content classifiers, false-positive-rate disparity on propensity-score-matched content, using human validation and hypothesis tests
• Designed and scaled ML fairness measurement across Google's policy classifiers, finding classification disparities affecting protected groups and enabling their fixes

Incremental spend from a partner incentive fund2020-2021

• Scoped, built, and delivered a self-updating dashboard measuring the daily incremental spend of Google's channel sales partners from a multi-million-dollar incentive fund, using counterfactual time-series forecasts fit on pre-fund spend

Measuring sales-program incrementality2020

• Wrote the requirements, built the data pipelines, and tech-led the team that shipped synthetic-control causal inference for 3 sales programs worth hundreds of millions of dollars where trials were not viable, validated on synthetic data and pre-treatment placebo tests

Advertiser churn prediction2019

• Built advertiser churn prediction end to end, from dataset and metrics to AutoML logistic-regression models and an automated lead pipeline and dashboard Google sales used to pick at-risk advertisers for high-touch outreach
• Sized a churn-prevention pilot for Google's at-risk advertisers with iterative power analyses, including sensitivity to the churn rate, and designed the framework assigning advertisers to test and control groups

Revenue-based routing for new advertisers

• Built the data and real-time pipelines for a team model that classified each new advertiser's 91-day revenue bucket to pick its sales treatment, routing $8B+ in revenue across 7M+ new advertisers
• Raised the precision of Google's new-advertiser revenue-routing model by redesigning its time-varying features as trend lines fit to each account's recent history, gaining 20%+ on high-spend advertisers and 28% on high-touch routing
• Mentored and onboarded new members of the team behind Google's $8B+ new-advertiser revenue-routing model, kept its data pipelines stable, wrote the model's documentation, and presented its results to leadership