Case study

From Instagram to engineers' daily workflow: bringing AI into Meta, 2021–2025

2021–22: used AI to make Instagram's Product Tagging for Everyone work. Late 2022: I proposed ChatGPT for developer tools weeks after it launched. 2023: I won approval for Meta's first use of GPT-4 on production code and proved it out with prototypes. 2024: I took it the last mile, with about 40% of my platform's hands-on engineers using AI coding tools. 2024–25: built AI into advertiser products.

Before ChatGPT: AI at Instagram

In 2021–22 I led Product Tagging for Everyone on Instagram, which let anyone, not just brands and creators, tag the products in their posts. The hard part was friction: asking people to search a huge product catalog to find the exact item in their photo would kill the feature. So I proposed applying AI to people's posts: finding visually similar products in the catalog and suggesting them at the moment of tagging. That became how the feature worked.

I led the launch across 8 teams in Ads and Instagram. It produced a $2.2M+ proven lift in sales (GMV), +77% weekly product-tag creation, and 100M+ product-tagged posts as of 2022. The full Social Commerce story.

The problem

In late 2022 I was the product manager for developer infrastructure in Meta's Reality Labs, the group building the Quest VR headsets and their operating system. We wanted to shrink the OS release cycle from about 10 weeks to 3. When we surveyed our engineers, their top request wasn't faster machines. They wanted better documentation, education and help tools.

The early bet

ChatGPT had just launched. I wrote an org-wide proposal to apply large language models like ChatGPT to developer documentation, education and help, and worked with two OS engineers to evaluate what the models could actually do on our code.

Winning permission

In 2023, using an outside model inside Meta meant clearing a long list of security, privacy, legal and policy reviews. I co-led Meta-wide discussions on LLMs, formed a cross-team working group in Reality Labs, and took the proposal through more than 20 review checkpoints and several escalations. The result was approval from Reality Labs leadership, including VP of VR Mark Rabkin and Meta CTO Andrew Bosworth, to use GPT-4 on Microsoft Azure in a sandboxed environment to help our OS developers write unit tests and code. It was Meta's first approved use of GPT-4 on production code.

Proving it works

I then formed and led an Applied AI working group for Reality Labs developer infrastructure. I wrote the product requirements and strategy, secured funding and dedicated engineering time from engineering leadership, and brought together the test authoring, test execution, internal AI and test-generation teams around one plan. I drove access to GPT-4 in VS Code, and ran proofs of concept for:

  • C++ unit-test generation. The generated tests compiled, added coverage and passed.
  • OS compatibility tests, with a goal of cutting test-writing effort by about 70%.
  • Visual app-compatibility checks with GPT-4 Turbo with Vision, reading screenshots from automated test runs to spot problems ranging from black screens to render tearing.

The last mile: daily use

Approval and working prototypes open the door, but the real test of a developer tool is whether engineers reach for it every day. In 2024 I led product for Meta's Ads Metrics Platform, about 60 engineers across four teams, counting senior ICs and engineering managers, building the measurement infrastructure that processes trillions of interactions for 4 billion people and 300M+ advertisers, businesses and creators. By then Meta's own AI coding assistant had matured. So I focused on habits, not permission. I showed engineers AI on their own day-to-day work (design docs, code generation, system design and code review) at the platform all-hands and the Metrics org's knowledge-sharing sessions, and ran recorded live sessions so anyone could learn at their own pace. I got 21 of the 50 hands-on engineers (42%) onto Meta's AI coding assistant for VS Code during early access. Engineering managers credited those sessions with starting to shape the org's habit of using AI tools for coding and writing.

AI in the product, not just the workflow (2024–25)

At the Ads Metrics Platform I also pushed AI into what we built for advertisers:

  • ML-modeled conversions. I drove the strategy and planning for a product-level reporting alpha, launched in 5 months, that gave 20 top advertisers access to Meta's ML-modeled conversions alongside their onsite and offsite metrics. 98.8% of report requests returned metrics without failing or timing out.
  • Real-time insights with forecasting. I spotted the opportunity to give top-tier advertisers real-time ads insights with forecasting, built consensus with the team and leadership, and ran the alpha test.
  • Anomaly detection and ML recommendations. I wrote the MVP proposal for automatically detecting outliers in ads metrics, and proposed building machine learning into Meta's metrics storage system to surface opportunities and recommendations for advertisers.

What I learned

Bringing AI into a large engineering org is two separate jobs. The first is winning permission: earning trust from security, legal and leadership, and narrowing a big vision to the use case that's most defensible and pays off soonest. In 2023, unit-test generation was that wedge. The second is winning habits: meeting engineers inside the tools they already use and showing them the payoff on their own work. Doing both is what turns an early bet into everyday practice.

Official redacted performance reviews available privately to clients with NDA on request.