William Gervasio

Member of Technical Staff @ Microsoft AI

Member of Technical Staff at Microsoft AI on Copilot, serving 100M+ monthly users and the youngest on the core product team. I own critical model behavior layers, architected response behavior for Copilot Tasks, and led evaluation harness work for fleets of coding agents. Before that at Instacart, I led receipt reconciliation work operating on 86% of orders, improving catalog linking from 50% to 90% across a $3M+/month workstream.

Previously built computer vision pipelines for cancer research at the Nobel laureate-founded Genome Sciences Centre and taught thousands of CS students as a Senior TA at UBC.

Member of Technical Staff - Core ProductMicrosoft

August 2025 - Present

One of five owners for every model feature for Consumer Copilot's 100M monthly active users, owning the release gate for prompt and evaluation changes, rewriting key product behaviors, and mentoring senior ICs on agentic features across releases.
Architected the response layer for Copilot Tasks, Microsoft's priority consumer agentic product, coordinating browser use, connectors, research, and document generation for long-running tasks with 40% retention and 200k daily active users under one month.
Designed and implemented backend integration architecture in C# enabling interoperability between two independent Copilot agent systems — auth flows, context routing, and event translation layers for cross-agent execution. Demo'd to Microsoft C-Suite.
Tech lead for the agentic harness evaluation platform profiling 1,000+ PRs per week from fleets of coding agents to deliver code improvements for Microsoft products serving 500M+ users.
Eliminated $2M+/month in Anthropic inference spend across Copilot by redesigning context management for every subagent.
Drove growth initiatives accounting for 50% of premium consumer feature usage, including one of four statistically significant November growth experiments with backend logic, prompts, and Python evaluation tooling.
Established early-career hiring standards by driving 300k+ social media views, sourcing 50% of intern candidates given an offer, calibrating the interview bar, and training interviewers for the organization's first early-career cohort.

MicrosoftCopilot Tasks: From Answers to Actions

CopilotMicrosoft Copilot

Software Engineer IIInstacart

July 2024 - August 2025

Platform lead for receipt reconciliation engineering and algorithms worth $3M+/month, improving receipt-item linking to online catalogs from 50% to 90% while operating on 86% of Instacart orders.
Redesigned the model training and serving lifecycle to improve training and inference latency by 20x, re-implementing offline ML and AI algorithms in PySpark while optimizing the real-time receipt understanding microservice with asynchrony and caching.
Owned internal APIs distributing Point-of-Sale data across the company, fueling fraud, in-store pricing, and shopper quality workstreams responsible for millions per month in savings.
Defined the evaluation platform from methodology to implementation for receipt item matching, fraud detection, and shopper mistake attribution to train real-time intervention models and boost human auditor velocity.
Maintained critical card payment flows (Stripe, Marqeta), fraudulent purchase rejection, and tax suppression governing every shopper transaction for 99.99% reliability across every Instacart retailer and shopper payout.
Reduced company-wide logging costs by $800k/month by implementing and enforcing log sampling across all Python services and team-specific Ruby services.
Promoted to Software Engineer II ahead of the earliest standard timeline, ranking in the top 5% of performers within the first six months.

Instacart ShoppersInstacart Shoppers – Get Paid to Shop

Instacart AppInstacart – Grocery Delivery

Cancer Researcher, Machine LearningBC Cancer

May 2023 - April 2024

Designed the end-to-end ML pipeline architecture for tracking 700+ cells dispensed per minute, linking physical and genetic cell features — a novel approach for a Nobel laureate-founded lab's cancer research.
Combined Bayesian inference, computer vision, and deep learning into a unified tracking system achieving 0.73 multi-object tracking accuracy.
Accelerated image labeling from months to minutes using motion-prompted foundation models on high-performance compute clusters, unblocking the broader research team to iterate faster.
Cut segmentation error rate by 50% with circular region proposals and data augmentation on Mask R-CNN (PyTorch).

Genome Sciences CentreCanada's Michael Smith Genome Sciences Centre

BC Cancer FoundationBC Cancer Foundation

Senior Teaching AssistantUBC

September 2020 - April 2024

Led programming labs and assessment across Software Engineering, Statistical Models, and Algorithms courses in R and Racket.
Designed and built the full autograding suite for Statistical Models, handling complex data outputs from ML algorithms.
Led delivery of UBC's Software Engineering MicroMasters on edX (188,700+ students); coordinated sprints, code reviews, and deadlines for full-stack projects in Racket and TypeScript.
Managed lecture and lab supervision, providing software design feedback and calibrating TA output quality.
Awarded Science Undergraduate Society Peer Helper Award for outstanding service and impact on peers.

CPSC 110CPSC 110 – Computation, Programs and Programming

CPSC 320CPSC 320 – Syllabus

Warrant Officer 1st Class: Fitness & SportsDND Canada

June 2018 - August 2020

Managed groups of 20+ cadets at a time while coordinating 3+ staff across barracks, drill, ceremonial, mental wellbeing, first aid, and morale. Coached cadets on how to lead others and conduct assessments.
Ranked in the top 13 nationally for leadership during regional staff placements; nominated best staff cadet in 2019.
Planned and led 18 weeks of sports and fitness training — organized tournaments, refereed competitions, and ran a tabloid event for 1,000+ cadets.
Commanded and coordinated 100 cadets on parade as Company Sergeant Major.