Sachin Gupta

Sachin Gupta is a member of technical staff at eBay and the creator of HeapLens, an open-source tool for diagnosing Java and Android memory problems inside Visual Studio Code. He applies the operational discipline of financial infrastructure to AI systems: making autonomous behavior controllable, generated code reviewable, and production data systems observable.

Gupta began his career at Tata Consultancy Services in 2011, contributing to an electronic shareholder-voting application for NSDL. He subsequently developed merchandising software for Walgreens and specialty-pharmacy systems supporting prescription tracking, auditing, reporting, and manufacturer-contract compliance. Since joining eBay in 2021, he has worked on its next-generation billing platform, including seller-and-balance migration, public APIs, and reconciliation pipelines. His more recent work encompasses AI-enabled production systems, agent governance, and retrieval infrastructure.

• HeapLens and private memory analysis. HeapLens combines a native Rust analysis engine, dominator trees, leak detection, source-code navigation, and HeapQL, a SQL-like language for querying heap objects. Its AI assistant can explain findings and generate executable queries; local-model support keeps sensitive heap data on the developer’s machine. HeapLens was selected for the SPLASH/ISSTA 2026 tool-demonstration program.

• Feature flags for autonomous agents. Gupta applies established release controls to agent behavior, governing prompts, tools, models, memory, autonomy, and sub-agents independently. His approach combines segmented rollouts, restricted tool authorization, adjustable memory retention, staged autonomy, and runtime kill switches. Controls must be reevaluated during active conversations and inherited by child agents; otherwise, an emergency shutdown can leave an autonomous process running.

• ReviewDebt and the cost of generated code. ReviewDebt estimates pull-request review burden using deterministic signals: change size, cross-file coupling, test evidence, ownership spread, AI-assistance indicators, and the author’s explanation. Gupta rejects opaque LLM-based judging in favor of reproducible scores and actionable reviewer guidance. Across an illustrative scan of 524 pull requests from three public repositories, structural complexity and submission volume mattered more than AI authorship alone.

• Production retrieval-augmented generation. Gupta treats retrieval systems as continuously maintained infrastructure, emphasizing incremental indexing, deletion across derived document fragments, embedding-model migrations, and traces explaining how answers were assembled. Reliable retrieval depends on synchronizing user-facing responses with the slower, ongoing maintenance of the underlying knowledge base.