top of page
droidcon+icon.png

20 NOVEMBER 2026

BENGALURU

DroidCon_icon_rotatet.png
Anita_Singh_edited.jpg

( SPEAKER )

Aman Kumar

Software Development Engineer II

Blinkit

Aman is an Android engineer at Blinkit, working on consumer-facing Android experiences at scale. He is passionate about Android performance, memory optimization, Jetpack Compose, and building reliable mobile experiences for millions of users. He enjoys solving complex production engineering problems and turning them into practical lessons for the Android community.

Session

Beyond Prompts: Building the Context Layer for AI-Driven Feature Development at Blinkit

What happens after the perfect prompt? An agent can write good code, but shipping a change to a production Android app also needs product context, the contracts between layers, a check on a real device, and evidence a reviewer can trust. At Blinkit, a feature rarely ends inside one Kotlin module. It can cross the Android client, server-driven UI, backend services, design tokens, analytics, and tests, spread across many repositories. Writing the Composable is the easy part. The hard part is helping the agent find everything that has to change, and showing that it hasn't broken a screen that already worked. In this session we follow one feature from a Figma design to a verified change: • Understanding context: What context means for AI-driven development, why source code alone is insufficient and how product and engineering knowledge changes the quality of the agent’s decisions. • Why more context can make things worse: how tokens and the context window work, why a giant CLAUDE.md backfires, and what context engineering means in practice. • Finding the right context: How structured knowledge about screens, flows, endpoints and existing components is combined with code search, semantic retrieval and graph traversal through MCP. • From design to plan: breaking down a Figma design, finding the closest existing component, and deciding where the change belongs before any code is written. • Building across layers: one repeatable workflow that plans, makes the smallest change, runs the tests and checks the result on a device, including the screens that must not change. • Measuring it: the same question asked with and without the context layer, compared on time, cost, tokens and whether the answer was right. This is not about an agent that magically understands the entire codebase. It is about reducing blind guesses, making review cycles smaller and giving the agent enough context to explain why each layer needs to change. By the end, you’ll have a practical blueprint for building AI-driven features that can work on real products at scale.
bottom of page