Shiv Bagavathy spent six years at Meta and is now building Turf, a social sports consumer app. He is known for rapidly prototyping and shipping projects, including rebuilding a browser assistant app in a single weekend. This episode walks through his actual development stack, how he manages multiple projects, and how he uses AI tools in practice.
Early AI experiments
In 2016, before the modern LLM era, Shiv built a chatbot called Wonder with a friend.
The idea was to help people remember things they often forget, like gate codes.
It used wit.ai for natural language understanding and TF-IDF-based information retrieval under the hood.
He describes it as a glimpse of what AI tools would later become.
Web browsers
Shiv uses three browsers daily:
Arc is his daily driver, even though it is no longer actively developed or maintained.
He values its spaces feature, which lets him separate work and personal profiles.
He also likes its distraction-free mode triggered by a simple shortcut.
Dia, an AI browser that lets users chat with their tabs.
He appreciates the chat-with-browser concept but has two main frustrations:
He wants more agent-style browser control, not just chat.
Dia does not preserve a searchable history of past conversations.
Chrome is his fallback for dev console work and testing.
He built his own browser extension that replicates Dia’s chat functionality while keeping all data local in the browser.
This extension lets him phase out Dia.
He notes that the company behind Dia reportedly raised around $40 million, and he essentially rebuilt its core value proposition in a weekend.
IDE and coding tools
Shiv has cycled through Cursor, Warp, Zed, and plain terminals, and is currently back to VS Code as his base editor.
He still runs Cloud Code inside Cursor for certain workflows.
His main complaint with AI coding tools is that they are not yet trustworthy enough to use hands-off.
He does not fully trust generated code and wants to review what the AI produces.
He believes less code means fewer bugs, so he wants agents to write the minimal amount of code necessary.
He uses a multi-agent review workflow:
He pits Codex, Cloud Code, and other agents against each other.
He has a sub-agent called “code review uncommitted” that reviews recently written or uncommitted code changes, especially those made by Codex or other automated tools.
He only steps in personally when the agents are circling without converging.
Task management
Shiv juggles a day job and many side projects.
For task management, he has tried many to-do apps but settled on Apple Reminders.
He uses it for everything from shopping lists to recurring tasks like changing bedsheets every two weeks.
He also keeps a physical paper checklist in front of his desk as an analog backup for daily priorities.
He uses a custom Codex skill called “today” that pulls his top three Linear tasks and any GitHub PRs needing attention.
Building “Shiv’s List”
Shiv built his own Craigslist-style marketplace app called Shiv’s List to sell personal items.
He chose to build it rather than post individually on Facebook Marketplace because it would not scale with volume.
The app lets users browse items, heart them, and claim dibs with a message.
People have actually used it during moves, and some discovered it through Facebook Marketplace without initially knowing Shiv built it.
He is now extending it into a two-sided marketplace:
One side is for things he wants to sell.
The other side, called “wants,” is for things he is looking to acquire, such as an Eames work desk priced around $1,800.
His build process for the new feature:
He starts with a rough sketch in a notebook or low-fidelity Figma.
He then moves directly into code, using Claude to spec out new components like a wants table and proposals table.
He uses a UI skill that constrains the LLM to proper CSS defaults and best practices for animations like Framer Motion.
When errors appear, he takes a screenshot and sends it back to the LLM with a simple “please fix” prompt.
He also has the LLM seed the database with sample wants so the feature is testable immediately.
He describes this workflow as closer to “one-shotting” a feature rather than iteratively re-prompting.
Shiv’s current stack
His preferred database is Convex.
He sees modern AI-assisted development as shifting computer science closer to an art form than pure engineering.
The limiting factor is no longer feasibility but imagination.
He acknowledges that some of the tools he uses today may be obsolete within weeks, but this is the stack he currently prefers.
