Chai Asawa

Chaitanya “Chai” Asawa leads engineering for AI clinical decision support at Abridge, a role he announced on January 5, 2026. His work centers on bringing patient context and clinical evidence into clinicians’ workflows. In a March 9 public statement, he described the goal as: “Not replacing judgment, but supporting it in real time. Connecting evidence directly in the medical record.”[2][1][6]

Asawa and Matt Troup, PA-C, coauthored “From Ambient AI to Clinician-Directed Intelligence”, published March 8, 2026 and subsequently modified August 27. They framed clinical documentation as an engineering problem beyond generating fluent text: output must be “contextual, steerable, and stable under refinement.” Their principle—“The note ultimately belongs to the clinician”—places responsibility for emphasis, interpretation and judgment with the person providing care.[3]

The article introduced natural-language prompt editing as a way for clinicians to direct a draft without manually rewriting it. Examples include making an assessment more concise, making a differential explicit, explaining why outpatient follow-up was chosen over imaging, and adding patient action items. Asawa and Troup connected this editing capability to decision support: a clinician can review cited evidence, decide whether it changes the plan, and then incorporate the reasoning into the note.[3]

They argued that decision support should be embedded, contextual, transparent and documentable: available where clinicians already work, grounded in the patient encounter, supported by inspectable citations, and easy to incorporate into documentation. They described an agent that retrieves trusted clinical content and uses the encounter to contextualize its results. Their account also described internal evaluation of precision, recall and faithfulness to source material.[3]

In his May 14, 2026 interview, Asawa compared Abridge with his previous work at Glean: “Context is what puts them to work.” Patient information, payer guidelines and medical literature can make models useful, but healthcare changes the consequences of error. He used prescribing something a patient is allergic to as an example of a potentially fatal mistake, explaining why clinical work requires greater rigor in offline evaluation and progressive rollout.[1]

Asawa framed assistance during a clinical conversation as a simultaneous quality, latency and cost problem. Prior authorization illustrates the difficulty: procedures, payer policies and health-system requirements create many combinations to resolve quickly. He raised intermediate representations of policies as a possible way to make that problem tractable. For latency more broadly, he discussed identifying bottlenecks, post-training models for efficiency, and using a cheap, fast model to triage work before handing it to a larger one. He acknowledged that improving models could make some present techniques temporary.[1]

His model strategy combined proprietary and third-party models. He described proprietary models as worthwhile when unique data enables higher quality, or comparable quality at lower cost and latency, citing transcription, diarization and note generation as examples. He also distinguished the real-time ambition from then-current behavior: in May 2026, support relied more on batching and text output, while prototypes explored triggering workflows at appropriate moments in a conversation. He regarded a spoken agent joining the encounter as potentially too disruptive at that stage.[1]

Asawa favored separating clinician preferences from rapidly changing underlying models, describing a separate memory store as an initial direction. Background agents or jobs could extract and consolidate useful preferences from clinician actions, note edits and transcripts. He also proposed treating the electronic health record conceptually as a filesystem: it holds more information than fits into a model’s context window, so agents that can read and manipulate data could help make that information usable.[1]

Healthcare also changed his release habits. Asawa recalled previously moving quickly from an alpha to general availability; at Abridge, he advocated early contact with actual use through progressive rollout, so feedback could improve the match between offline evaluations and real-world conditions. He described transcript de-identification models as another prerequisite for learning from clinical data, emphasizing that this probabilistic step itself needs validation before subsequent training or evaluation, alongside appropriate partner contracts.[1]

Before Abridge, Asawa worked in enterprise search at Glean; that experience supplies the background for his comparisons between organizational and clinical context. In a January 20, 2026 account, Annelies Gamble, who disclosed their friendship, described him as a founding engineer who had left Glean after more than six years. Earlier still, he worked in robotics and AI research at Vicarious AI, including work on intelligent agents, as recorded in his Glean author biography.[5][4][1]

In his conversation with Gamble before leaving Glean, Asawa explained that short workplace queries require organizational signals to reveal intent: the answer can depend on role, team, location and recent activity. He argued that horizontal platforms need deliberate investments in particular domains to meet large enterprises’ requirements. Approvals, permissions, auditing and escalation were examples of shared workflow primitives: building them deeply in one domain could improve the product elsewhere. In his later Abridge interview, he credited Glean’s durable foundations to high-quality search and colleagues with experience building search at Google.[5][1]

His interest in healthcare AI predates the Abridge move. In his January 2026 joining announcement, Asawa recalled helping start an AI-in-healthcare seminar more than eight years earlier, motivated by technology such as AI skin-cancer detection and problems of access and cost. What stayed with him was how little of that technology reached practice. He presented Abridge as an opportunity to pursue clinical decisions amid improving technology, clinician adoption and regulatory tailwinds.[2]

Asawa expects stronger agents to increase the importance of context layers and tools, while reducing the need for scaffolding that compensates for limited model capabilities. He identified event-driven technology—Kafka, Temporal and sockets—and collaboration patterns such as CRDTs as potentially durable foundations for agent systems. He also resisted treating prototypes as a universal replacement for written specifications: “Judgment and clarity always matters.” His view allows quick prototypes for small changes while retaining written clarity for complex work.[1]