$ Hamza Harkous

Teaching AI models requires massive amounts of high-quality data, but generating it manually is impossible.

Simple automated prompting causes AI to create repetitive, low-quality data. I co-founded Simula with Benoit Seguin to solve this. We designed a programmable framework that independently controls data diversity, complexity, and quality without human intervention.

Our team scaled Simula into Google's primary internal multimodal data synthesis engine. It is now used by over 2,500 Googlers, was used to generate trillions of tokens, and achieved a 93% user satisfaction score, the highest among all data tools at Google. Used by hundreds of teams, today it is a key enabler for the Gemma ecosystem, provides the primary synthetic data backbone for Gemini safety classifiers, powers production user protection features like AI-powered scam detection, and extends much further across frontier AI applications and data use cases.

Mobile apps often promise one thing in their privacy policy but do something completely different in their code.

Auditing this manually is impossible. I led the ML strategy and built core AI models for privacy policy analysis that were fundamental to launching Google Checks (an Area 120 project). I architected the end-to-end ML pipeline: starting from continued in-domain pre-training on over 1 million policies, training high-capacity T5 11B teacher models, building custom datasets, and distilling them into significantly smaller models.

This ML architecture scaled the system for production serving, allowing it to recognize 70+ legal compliance tags with high precision. Checks launched globally as Google's automated compliance platform, helping developers verify their apps before release.

Users write millions of app reviews daily, but finding genuine complaints in that unstructured noise is like finding a needle in a haystack.

In 2021, we were building specialized LLMs for user feedback. I designed a deep learning pipeline that reads and understands colloquial user complaints, mapping them directly into formal categories.

I scaled Hark, launching Hark's Model Service to power feedback triaging across Google. This became a core analysis engine processing hundreds of thousands of messages and reducing triaging cycle time by 2X. I also drove efficiency gains through distillation, distilling to ~20x smaller models to meet strict QPS requirements. We published the core research at IEEE S&P 2022.

Turning structured data into natural language without hallucinating facts was a persistent challenge in conversational AI.

In 2019, neural text generation was prone to fabricating facts, and production conversational systems still relied heavily on rigid, hand-crafted templates. At Amazon, I created and led DataTuner to mitigate this hallucination problem end-to-end. We decoupled fluency from factual accuracy: fine-tuning a GPT-2 model to generate diverse, fluent candidate descriptions directly from structured records, paired with a specialized RoBERTa semantic fidelity classifier that audited each candidate to catch omissions or hallucinated details.

By separating generation from semantic verification, we achieved state-of-the-art semantic accuracy across standard data-to-text benchmarks (WebNLG and Cleaned E2E) without relying on dataset-specific heuristics. We published the system at COLING 2020 and open-sourced the framework, which was featured by Amazon Science for its principled approach to faithful text generation.

Privacy policies are notoriously long, vague, and filled with legalese that nobody reads (and still are).

During my PhD at EPFL in 2017, we helped democratize the use of deep learning for privacy policies, building the most successful and widely adopted system at the time. I built classifiers that segment complex legal text into standardized practices, and created PriBot, an AI chatbot that answers user questions about privacy policies.

I launched Polisis and PriBot as public web services, and the technology was later licensed to major companies, including DuckDuckGo. Over 45,000 users used the tools to inspect privacy agreements. The project was featured in Wired and The Wall Street Journal for making online privacy accessible.