Palak Agarwal leads developer relations at Reducto and created Serve, an AI photography coach for iOS that offers guidance while someone can still change the shot. Her work spans enterprise identity software, consumer AI, and explaining how document-processing systems turn complex pages into usable data.
From enterprise identity to photography
Palak studied at the University of Waterloo and worked in product roles at Microsoft, Gamma, Scotiabank, and Arteria AI. That work included collaborating with engineers, designing user studies, and building internal evaluation frameworks.
During her summer 2025 AI product-management internship on Microsoft Entra, customer calls revealed difficulties with self-service account capabilities that created additional support work for IT administrators. Palak independently scoped, designed, and engineered an agent prototype to address the problem, then planned further iteration with customers and product managers. Her account of the internship also weighs the tradeoffs between large companies and startups: experienced colleagues and organizational structure at a large company, alongside opportunities for faster experiments and broader responsibility at a startup.
Her experience at Gamma helped draw her toward consumer AI and products people could use intuitively. During a gap term from Waterloo spent in San Francisco, she began building Serve. The app was designed to provide real-time photography coaching on composition, lighting, and framing, personalized to the photographer’s style. Offering guidance during capture gives the photographer a chance to adjust the scene or camera position before taking the picture.
Making document AI understandable
At Reducto, Palak’s technical communication includes editing the team’s research on evaluating figure summaries, written by Chuyi Shang and the Reducto ML team. The research addresses a problem that simple transcription checks cannot resolve: two descriptions of a chart can both be accurate while preserving very different amounts of useful information.
Counting errors alone could reward summaries that omitted nearly everything. Rewarding length introduced a different distortion: a logo may need only a few words, while a complicated diagram requires several sentences. The team combined diagnostic error scoring with pairwise evaluation against a fixed baseline to assess summaries. Palak edited the explanation of that work; the research and evaluation methods belong to Chuyi and the ML team.
Palak also introduced the document-processing layer in a joint presentation with Omar Alhait on the Epstein files and JMail. Their case study connected handwritten logs and heavily redacted PDFs to the structured records needed for searchable applications. Omar demonstrated the extraction pipeline behind JFlights. The distinction matters to understanding the work: parsing must preserve handwriting, layout, and reading order before a schema can organize the extracted information into useful fields.
Supporting early founders
Alongside her product and developer-relations work, Palak invests in pre-seed teams through the a16z speedrun scout fund. Her earlier work with Afore included finding student founders, training scouts, and organizing founder gatherings. Her internship essay also describes hosting a student gathering with Afore, extending that work into opportunities for prospective founders and collaborators to meet.
Palak Agarwal and Omar Alhait explain how Reducto turned difficult PDFs and handwritten flight logs into data for JMail—and why schemas, redactions, document-specific settings and representative evaluations matter before scaling.
JFlights needed a flight-specific extraction configuration that could handle varied handwritten and redacted documents under one schema.
▶ Watch ↗