Kabir Goel

Kabir Goel leads product engineering at Cartesia. His technical writing addresses two practical challenges in AI products: delivering generated audio quickly enough for real-time use, and giving coding agents the feedback they need to judge whether their changes work. He graduated from UC Berkeley in 2024, where he studied computer science and public policy.[2][5][6]

In a February 6, 2026 Cartesia essay, Goel framed the distinction this way: “Writing code feels solved. But engineering is not.” In his account, engineering includes judging a feature’s impact, responding to real-world feedback, and continuing until the company’s goals are met. He described Cartesia’s adoption of coding agents as organic experimentation rather than a top-down mandate, with multiple review agents producing sometimes complementary and sometimes contradictory feedback.[6]

Goel reported that the team had used Opus 4.5 to safely remove a feature exceeding 10,000 lines, while still struggling to get agents to assess production traction or connect an API latency spike to code. His diagnosis was broken feedback loops: an agent might write an A/B test without being able to inspect PostHog results, deploy a UI fix without seeing its Vercel preview, or optimize a query without checking production metrics. He advocated safely providing access to repositories, CI, previews, dashboards, and analytics so agents could verify outcomes and iterate.[6]

He treated that infrastructure as a dedicated engineering function. Tracking new tools, experimenting, and turning useful findings into internal systems, he argued, required sustained attention. The essay announced hiring for an AI & Developer Acceleration engineer to design developer experience for both humans and AI; faster CI and easier local setup would also benefit human engineers.[6]

In his March 10, 2025 post on first-byte latency, Goel explained why fast voice generation also requires attention to audio delivery. Drawing on Cartesia’s team experience, he recommended caching CORS preflight responses with Access-Control-Max-Age to avoid repeated round trips, and avoiding response compression on sufficiently fast connections because compression can buffer data before transmitting it. For constrained bandwidth, he discussed smaller audio payloads and Opus in a WebM container. Lowering the sample rate can affect sound, while disabling FFmpeg input buffering can introduce stuttering.[5][2]

He also examined the tradeoff between uninterrupted playback and low latency. A browser’s audio element automatically buffers for smooth playback, limiting the developer’s control over that tradeoff. Using the Web Audio API with fetched ReadableStreams offers control over chunk scheduling and buffering, but shifts complexity to the application: developers must handle timing differences and artifacts such as popping, potentially by finding zero-crossings in the audio data.[5]

Before those Cartesia essays, Goel explored text interfaces and built community software while at Berkeley. These earlier projects provide concrete examples of his interface and implementation decisions.[3][7][2]

On April 19, 2023, Goel delivered three LLM text-crafting demos at AI/UX: Beyond The Textbox, hosted at Notion HQ. Latent Space’s event recap lists a projectional editor for text, a tone-switching palette inspired by FigJam, and zoom previews for text.[3][1]

In his July 14, 2023 account of Little Boy Blue, Goel described replacing Berkeley Blueprint’s unreliable Slack kudos bot after recently serving as the student organization’s VP of Projects. Members awarded points by mentioning someone with ++; crashes in the old bot meant recognition could be lost. Goel built the replacement on Replit using Slack’s Bolt library for Python, a key-value score store, and rule-based message parsing. He blocked self-awards, added cooldowns for repeated awards, and deduplicated mentions within messages.[7][2]

He replaced the old leaderboard’s custom message syntax with a discoverable slash command and made results private to the requester except in the dedicated leaderboard channel. To preserve existing scores, he indexed the details of a few hundred Slack users and mapped old usernames to user IDs. Threaded replies and clearer formatting reduced channel clutter. Bry drew the bot’s mascot and its context-specific variants. Goel reported that the move from AWS to Replit halved Blueprint’s hosting spending and that, by the July 2023 post, Little Boy Blue had awarded more than 4,000 kudos.[7][2]