Jeffrey Wang

Jeffrey Wang is an Amplitude cofounder and former chief architect. His work connects the infrastructure for analyzing user behavior with AI features that help people query that data. In a June 2023 discussion with Joe Reeve, he described an LLM-powered copywriting prototype he built, explained the team's move from general-model calls to embeddings for event selection, and argued that AI features should be judged by customer value and business outcomes.[3][1]

At an internal hackathon roughly two months before the 2023 discussion, Wang built a prototype that continually generated copy variations with an LLM and used multi-armed-bandit selection to identify the best variants. He saw it as a small example of self-improving products: a person supplies the product's purpose and a structure the model can modify, while generated alternatives and experiments help refine titles, descriptions, or navigation. This was a prototype and a proposed direction, rather than a released capability.[1]

Wang advised builders to connect LLM features to the outcomes they already measured, such as conversions and customer value. His preferred north-star metrics correlated with business outcomes. Time in an app could be useful for some social products, he said, but longer search sessions could indicate failure; optimizing an inappropriate metric could make a product worse.[1]

Wang viewed natural language as an accessible starting point for people unfamiliar with a dataset or unsure which questions to ask. He also emphasized its ambiguity when an analysis needs precise semantics, such as calculating revenue. As he put it in the 2023 interview, “It's actually easier to make SQL or code precise than it is natural language.”[1]

For chart generation, Wang advocated breaking the task into selections—chart type, metric, events, and properties—and providing examples specific to each task. Reeve, who ran Amplitude's AI R&D efforts, described the implementation as seven subquestions, with monitoring for both individual inferences and the complete request. Wang's explanation stressed why this decomposition helped: a general model lacked knowledge of Amplitude's domain, and generating an entire chart definition at once demanded too much from a single prompt.[1]

Wang also explained a collective decision about selecting the right event to analyze. The team first used a general OpenAI model to match a user's intent to event types, validating the idea quickly. It then invested in precomputed embeddings for repeated selection, which he described as faster and cheaper while retaining reasonable quality. In June 2023, he clarified that the PostgreSQL vector workload contained taxonomies—event types such as add to cart or purchase—with cardinality he estimated in the tens of millions. It did not contain individual end-user or event vectors; he distinguished those from a prospective workload involving trillions of events.[1]

In a March 4, 2026 guest post written in partnership with AWS, Wang described the team's subsequent retrieval architecture for Ask Amplitude. The account traces PostgreSQL cosine-similarity search through pgvector approximate-nearest-neighbor indexes to OpenSearch, which combined keyword and semantic search. Consolidating retrieval reduced synchronization pipelines to one and moved result merging, weighting, and ranking out of the application. The post does not date the migrations.[6]

The same account explains how retrieved context supports analysis: a series of LLM prompts converts a question into JSON for Amplitude's custom query engine. It reports narrowing schemas of 5,000–20,000 elements to 20–50 relevant elements per question, and extending retrieval to 20 million user-created charts and dashboards. Existing analyses could supply meaning absent from column names—for example, showing how colleagues had filtered data to answer a question about two-day shipping.[6]

In 2023, Wang argued that distinctive datasets offered substantial untapped potential for AI applications. He hoped models trained on behavioral sequences could improve predictions of purchases or churn, group users, and help explain differences in behavior. He treated causal interpretation as a difficult prospect, and was explicitly uncertain whether such models could make noisy, automatically collected activity useful enough to remove manual instrumentation.[1]

Wang described pseudonymized first-party behavioral data as a balance between collecting no information and tracking people across the internet. Identifiers made it possible to connect an earlier visit with a later purchase. In explaining Amplitude's philosophy, he said the company did not sell the data or mix it with cross-site tracking data, and distinguished data for improving a product from data for advertising.[1]

Wang's analytics infrastructure work predates these LLM experiments. He studied computer science at Stanford and gained infrastructure experience at Palantir and Sumo Logic. In his retrospective account of Amplitude's founding, his collaborators Spencer and Curtis had developed analytics tools while working on Sonalight, a voice-texting product they took through Y Combinator. Demand for those tools led them toward analytics. Wang recalled meeting them through mutual friends and joining as a distributed-systems and analytics specialist.[3][1]

In his May 25, 2016 article on Nova, Wang explained the team's decision to build an in-house column store. The earlier architecture's pre-aggregation made supported queries fast, but limited flexibility and became fragmented and costly as behavioral analyses grew more complex. Nova aimed to support those analyses while scaling cost-effectively. Wang acknowledged the engineering and reliability costs of building a database, and identified analytics and distributed systems as the team's core competencies.[5]

Nova organized computation around users, moving code to data so workers could analyze disjoint sets of users in parallel. For a shopping funnel, each worker could examine a user's registration, item views, cart additions, and purchases, then combine small results across workers. Events remained immutable; when two identities were later recognized as the same person, a query-time mapping reconciled them rather than rewriting stored events.[5]

The architecture separated real-time ingestion from batch processing while exposing a common columnar format to queries. S3 served as canonical storage, with other layers acting as caches over immutable chunks. Eventual consistency allowed a short delay between accepting events and making them queryable. Wang credited Google's PowerDrill, Facebook's SCUBA, Microsoft's Trill, and Metamarkets' Druid as design influences. His 2016 account reported processing more than 70 billion events monthly, regularly scanning 10–20 billion events per query, and completing 95% of queries in under three seconds.[5]