Project details

Role: Product Designer & Builder — concept, interaction design, and code (React/TypeScript)

Team: Cross-functional hackathon team — design, engineering, data science

Timeline: AI Hackathon 2025 → internal adoption path

Impact: Cut recurring leadership reporting time from multiple days per cycle to near-zero

Scope: End-to-end — from problem framing and agent architecture to interaction design and front-end code

Recognition: Winner, internal Uber AI hackathon

・・・・・・・・・・・

Executive summary

Leadership reporting at scale was a manual pipeline: days of querying, exporting, summarizing, and formatting — every cycle, across multiple teams. Insights arrived too late to act on.

I designed and built an agentic system that automates the entire path from raw operational data to executive decision: specialized AI agents retrieve the data, interpret it in marketplace context, assemble a consistent report, and connect every degraded metric to its verified owner — delivered where leaders already work.

The design core wasn't the interface. It was modeling how executives scan, interpret, and act under time pressure — and shaping what the agents produce around that mental model.

Reporting time went from multiple days per cycle to near-zero. The project won Uber's internal AI hackathon and became a reusable foundation for other GenAI-driven automations.

Instead of shipping a dashboard, I built a decision system.

・・・・・・・・・・・

The Problem:

Insights arrived too late to act on

• Dashboards with access and usability barriers, where insights had to be hunted, not discovered

• No scalable way for domain owners to detect problems and contextualize them with metrics

• No regular review cadence — the friction was too high

• Similar manual reporting pipelines duplicated across teams

The challenge

How might we redesign the path from raw operational data to executive decision — removing the human bottleneck without removing human judgment?

・・・・・・・・・・・

An agentic reporting system

I designed and built an end-to-end automation that replaces the manual query → export → summarize → email loop with a chain of specialized AI agents, each owning one step of the pipeline:

1. Data retrieval agent — orchestrates queries against the internal data stack automatically, on cadence.

2. Interpretation agent — summarizes results grounded in marketplace context and historical knowledge, not generic text generation.

3. Report design agent — structures insights into a consistent, scalable HTML report. Same hierarchy, same visual language, every cycle — accuracy and consistency without manual formatting.

4. Dashboard integration agent — one tap in the report opens the internal analytics tool pre-configured: visualization selected, markets pre-filtered. Leaders jump straight to insights, zero configuration.

5. Ownership agent — when a metric degrades, the system identifies the verified owner automatically and surfaces them in the report, one click away from a Slack conversation.

Delivery happens where leaders already are — email and Slack — without human intervention.

・・・・・・・・・・・

Designing for decisions under time pressure

The design work wasn't the interface polish — it was modeling how leaders actually scan, interpret, and act on information when time is scarce, and letting that mental model shape the system:

• Signal over noise: executive summary first, market health next, degradation alerts with impact — in the order a decision actually gets made.

• Every alert is actionable: a degraded metric ships with its impact, a suggested course of action, and a verified owner. No dead-end information.

• Zero-configuration depth: detail is one click away, already filtered. The system does the setup work a human used to do.

I prototyped and built the front end in React and TypeScript — the same AI-augmented, code-based workflow I use in my daily design practice.

・・・・・・・・・・・

Results

• Recurring reporting time: from multiple days per cycle to near-zero

• Consistent report structure across cycles and teams — zero manual formatting

• Faster path from degradation signal to accountable owner

• A reusable agentic foundation adopted as a reference for other GenAI-driven operational automations internally

• Winner of the internal Uber AI hackathon

・・・・・・・・・・・

Key Learnings

• AI systems earn trust the same way people do: by showing their reasoning and their sources.

• The highest-leverage design decision in an agentic system is what each agent owns — boundaries are the interface.

• Automation shouldn't remove human judgment; it should remove everything standing between judgment and action.

• Designers who build ship faster feedback loops: coding the prototype collapsed the distance between design intent and working system.

・・・・・・・・・・・

🔒 Additional Details

This system integrates with internal data infrastructure and tooling; specific tool names, metrics, and flows are redacted from this public version. A deeper walkthrough — including the agent architecture and demo — is available upon request.