Swaroop Chitlur Haridas

Swaroop Chitlur Haridas is a staff engineer at DoorDash working on machine-learning and generative-AI infrastructure, and the author of A Byte of Python. His work helps people use complicated technology: beginners learning to program, data scientists building production features, and product specialists deciding whether an AI system is good enough to ship.

From teaching programming to technical leadership

Writing under the name Swaroop Chitlur, he created A Byte of Python as a free introduction for readers with little computing experience. Its Creative Commons license permits sharing and adaptation, including translations; the open-source book repository gives readers a way to propose corrections and contribute improvements. He also wrote A Byte of Vim, extending his teaching to the programmable text editor.

His career has included large companies, consulting, startups, and building a company with friends. He helped found IonLab and left in November 2009. The experience sharpened his distinction between understanding advice and learning through execution: reading about entrepreneurship had not prevented the team from making familiar mistakes. He also emphasized empathy for customers whose motivations and working conditions can be difficult to understand from outside.

By 2019, Chitlur was an engineering manager at Helpshift. He subsequently returned to an individual-contributor role, explaining in his 2021 essay Back to IC that he wanted more hands-on involvement in the projects his team was doing. He treated management and technical leadership as paths he could move between as the work required. One challenge he recognized in himself was learning to identify the right problems rather than defaulting to solving the ones already placed before him.

Making machine learning easier to put into production

At DoorDash, Chitlur contributed to the team behind Riviera, a declarative framework for real-time feature engineering. Riviera lets practitioners express transformations through SQL and configuration while shared infrastructure handles much of the stream-processing setup. A feature such as a restaurant’s order count over the preceding half-hour can become a configured pipeline rather than a separately maintained Flink application. His involvement was part of a broader team effort, rather than sole ownership of the framework.

By 2023, he was also working with colleagues on Project Lucent, DoorDash’s machine-learning training platform. That work explored Ray and KubeRay to address growing training and forecasting workloads, infrastructure costs, development speed, and debugging. The progression from streaming features to training infrastructure and generative-AI services has kept his attention on the machinery that allows other teams to build and improve products.

His more recent evaluation work addresses a different bottleneck: useful judgment often belongs to people the engineering tools do not serve. Chitlur argues that product managers, strategy and operations teams, and domain experts need direct ways to shape AI quality. In his account of DoorDash’s evaluation work, he describes judgment becoming stranded in spreadsheets and email while teams rebuilt evaluation infrastructure independently.

Distinctive contributions and priorities

• Cross-functional evaluation. Chitlur favors a quality process that incorporates the people who understand the domain. A shopping assistant, a personalization system, and a multi-agent operations workflow require different judgments. Their shared platform must accommodate those differences while connecting human review to repeatable evaluation.

• Stable APIs, adaptable workflows. In his joint work with Nachiket Paranjape, the platform supplies common interfaces for traces, scores, datasets, and judges, while partner teams can build annotation tools suited to their tasks. Coding agents help operators create those interfaces without waiting for the central team to implement every variation. The division preserves shared infrastructure while letting the people doing the work shape its presentation.

• Reviewable judge calibration. The same joint work connects sampled traces, human annotations, trusted datasets, and calibrated model judges in a continuing improvement loop. Partners can inspect original and revised judge prompts rather than accepting an opaque optimization result. Prompt ownership can remain with product, operations, or engineering according to each team’s needs; the platform supports that organizational flexibility.

• Accuracy, latency and cost together. Chitlur’s platform work treats these as competing requirements for both models and agents. His current interests include moving real product workloads onto open-weights models, with evaluation, prompt tuning, benchmarking, and inference considered together. The practical objective is a production system that meets the product’s quality requirements while becoming faster and cheaper to operate.

Across his books, career reflections, and platform work, Chitlur repeatedly makes room for other people to participate. The beginner needs an explanation that starts at the beginning; the practitioner needs infrastructure that removes repetitive setup; the domain expert needs a workflow through which judgment can change the product.