meta trust & safety
technical program manager
Meta Cross Integrity > Content & Interactions Solutions. London (2022 - 2023)
the problem
Back in Oct 2022, while working in Integrity for the End to End Encryption messages of Facebook and Instagram, I realised how Scams were increasing and escalated my concerns.
After realising that the existing human review flow had poor consistency, that led to low precision on the classifiers for Facebook and Instagram, I convinced my senior stakeholders to fund an Engineering team for Scams.
This was important to catch bad actors but also reduce the bad user experience that over-enforcing was causing (there was a high volume of False Positives).
The goal, within one quarter, was to create a new human review policy and protocol, and models, to achieve 95% consistency and 90% of precision in the new classifiers.
This required a strong Technical Program Manager to lead all the efforts between Engineering, Policy, Privacy, Product, Operations, Intelligence and Legal. I was appointed to be the TPM for Scams.
the contribution
OpEx reduction of 12%. Increase volume enforcement of 192x. 50% less appeals
* Managed to convince senior stakeholders to prioritise an engineering team for Scams, backing up my conversations with data.
* We achieve the launch of a new human review workflow, partially automated, that it is now the baseline for data training and ground truth for improving classifiers precision. We can now cover +25% more tickets (reports), consistency increased to +95%.
* Delivered on time a classifier with 92% precision (in one quarter).
* As a consequence, appeals were reduced by half.
* Daily callibration calls for definition of Scams archetypes with the SMEs.
* Escalations resolutions of reported cases that are borderline.
* Implemented new process for guardailing the proliferation of ad hoc tactical solutions. Quarterly audit of enforcement, implementation of new standards and processes, with a new stress testing of the solutions before implementation.
* Hands on reduction of legacy pipelines: 95% reduction of machine capacity consumption.
* Avoided CMA litigation and consequent fine.
the learnings
Scams are a very adversarial space and changes at a high speed, so frequent auditing of solutions is a must.
Integrity solutions that require Machine Learning classifiers have to compromise between precision to reduce false positives or recall to reduce false negatives. There are certain levels of trade-offs.
Shifting the Operating Model to a new one, requires conversations, definition of roles. Sign off from senior managers before the implementation is a way to ensure that the transition is backed up.
In such a big organisation, you need to plan with high level of detail, get to know your stakeholders and build trust (Policy, Legal, Operations managers to seek human reviewers capacity…)