How to Build AI You Can Trust, With Adyen's Dhruv Ghulati
In a Nutshell
Dhruv Ghulati emphasizes building trustworthy AI through deep fraud pattern understanding, human-in-the-loop systems, and governance frameworks that balance performance with explainability. Key infrastructure investments include sandbox environments and democratized data access to enable rapid experimentation rather than pre-planned use cases. The core message: invest in accessible AI tooling and feedback loops while maintaining interpretability safeguards, as the current era rewards hands-on creative building over rigid planning.
These notes were generated by AI and may contain inaccuracies.
Dhruv Ghulati started his career in finance and banking, with his first job at Merrill Lynch on the trading floor. Around 2014, he became interested in technology and entered AI. His career path included being an early-stage founder in London, working on a company that detected misinformation and disinformation online. After exiting his first company, he focused on developing AI product management skills, recognizing that understanding product is essential for technical leadership and entrepreneurship. He moved to Amsterdam where he led the Uber AI team on product as the first site lead, working on AI problems across mobility, Uber Eats, delivery, and forecasting.
Dhruv currently serves as a principal product manager at Adyen, leading AI product development for payments. His career has consistently focused on detecting fraud and negative content, from identifying misinformation in his startup to detecting counterfeits and forged documents at Onfido, and now working on chargeback fraud detection at Adyen.
Effective fraud detection requires understanding how fraudulent information is generated. This involves going deep into fraud patterns, including how generative AI creates disinformation, the motivation of attackers and creators, and the techniques they deploy. Without this understanding, building effective detection systems is impossible. The process operates as a continuous cat-and-mouse game, with waves of creation followed by regulation, moderation, and safety measures.
The European AI Act has been a significant development in AI regulation, designed to catch up with the rapid pace of AI advancement. The goal is to avoid regulatory situations where rules slow progress unnecessarily or lack practical sense. Key regulatory considerations for AI include what data models are trained on, whether training data contains personally identifiable information, and whether models contain biases in their training data.
AI regulation focuses heavily on models making real decisions that affect people's lives, such as loan application approvals or denials. Previously, manual processes existed for reasons like requiring human expertise in loan or mortgage applications. Regulators are now questioning whether human-in-the-loop systems are necessary due to required expertise or whether machines could make decisions more fairly.
Companies range from highly locked-down environments with closed user groups, user permissions, and limitations on downloading code or files, to early-stage startups where everyone has access to everything. Access controls include permissions on data, code, and tools like Kafka pipelines or production databases. The key question is identifying the biggest risks of giving access versus the costs of restricting it, including the risk of incorrect data interpretation leading to poor outcomes.
Agents have potential to create permission layers automatically, reducing the need for manual security expert intervention and internal meetings to determine access rights. This parallels challenges in startups where founders are the only ones who can fix production issues, and similar permission problems emerge when companies onboard thousands of new developer agents that need access to production databases.
AI-native companies are creating MCPs (likely meaning Model Context Protocol or similar tool access systems) to provide access to different tools. This enables product managers to edit tickets directly from their terminal, adjust statuses, and make bulk changes—tasks that previously required manual point-and-click operations. Every SaaS company needs an MCP strategy to unlock access for new user bases, similar to how APIs promote open ecosystems.
Uber built the Michelangelo platform to support their marketplace real-time algorithm that matches drivers and riders in real time. This required running models at scale with low latency, constant pinging across different time zones, system redundancy, and continuous observability. Michelangelo enabled other teams to reuse components for real-time pricing, incentives, demand and supply prediction without rebuilding feature computation, MLflow, Airflow, or DAG pipelines repeatedly.
The MLOps role involves language around features, feature stores, model pipelines, observability, model evaluation metrics, retraining systems, and AutoML—distinct from data engineering or data catalogs. Large companies often hire data scientists to build models without understanding that model builders cannot succeed without architecture for running models effectively, running experiments quickly, testing offline and online, and versioning test datasets.
Traditional ML metrics include precision and recall, which can be tuned through ROC curves. In fraud detection, precision measures how often flagged fraud is actually fraud, while recall measures what percentage of actual fraud was caught. In AI systems like chat assistants, evaluation metrics differ and include conversation quality assessment, multi-turn versus one-shot conversation ratios, average conversation length, and determining what constitutes a good response.
RLHF (Reinforcement Learning from Human Feedback) is used to tune model personality characteristics like friendliness levels across GPT versions. While internal model metrics and open-source benchmarks exist for specific tasks, the actual online outcome that matters to management is whether chat GPT is growing, which may not correlate with making conversations friendlier. Growth segments may prefer targeted, rationalized conversations over friendly ones, or may need better spreadsheet access through tool calling.
AI governance requires access controls, data catalogs documenting what data exists and what jobs are running, and logging to understand data contents. When building models from historical loan application data that includes race or country features, safeguards like red teaming are needed to check if models are learning inappropriately from these features. The challenge is that models may accurately reflect historical biases in loan approvals, indicating problems in the underlying system rather than the model itself.
Removing problematic features may worsen model performance for the core task, representing a business decision to prioritize fairness over accuracy. Correlation analysis may reveal that models still correlate with removed features even when not explicitly using them. Governance and decision driver analysis can improve not just automated systems but also rearchitect underlying processes like loan approval or fraud detection.
In chargeback fraud detection, some features may be highly predictive but lack explainable sense. Companies face decisions between building the most precise and highest-recall systems versus sacrificing some performance to make systems more explainable and interpretable. Interpretability helps improve the overall system beyond just automation.
Rather than trying to map all AI use cases in advance, which will change quarterly, CTOs should invest in sandbox environments that enable anyone to build machine learning models. This includes opening up data for feature access, whether from business data streams, sales calls, or payment transactions with anonymized shopper information.
Open up anonymized transaction data from payments going through systems to enable internal teams to train models using simple approaches like XGBoost and logistic regression. Make it easy for people to experiment with template-based simple models and observe what use cases emerge naturally from this hands-on exploration.
Open access to LLMs alongside traditional models and establish clear channels where use cases and issues that arise during experimentation flow back to the central team. This feedback loop enables continuous patching and fixing while providing deep insights into what people actually want to use AI for within the organization.
This represents the age of building and creation, presenting significant opportunities for both workplace professionals and humanity broadly. The current moment offers a chance to recreate the magic of creativity across all roles, whether someone works in a traditional job or operates as an entrepreneur.
Weave creativity and agency into daily work by being hands-on with AI tools and approaches. Bring creative experimentation into regular workflows regardless of job function or organizational structure.
"It's the age of building. It's the age of creation."
"Be creative, have an agency, and be hands-on."
For additional information about AI integration within organizations, resources are available through provided links.
Keep JetBrains in your library
Save the videos and channels worth coming back to, and find them again in one place.
