Case study · 05
InsightFlow AI — autonomous report-to-action agent
Overview
Mobile-first executive copilot that turns messy business reports into one prioritized decision — multi-agent Gemini analysis, leadership alerts, consequence paths, and simulated Slack/email/CRM execution, deployed on web via Expo.
Project preview
Technologies used
- Expo
- React Native
- TypeScript
- Google Gemini
- Expo Router
- React Native Reanimated
Key features
- Five-agent pipeline (Reader → Main Points → Problems → Next Steps → Results) with live trace on Analysis
- Quick mode: single Gemini pass (~20s) with the same structured output for demos
- Document ingest via paste, .txt upload, or PDF text extraction through Gemini
- Executive dashboard: leadership alerts, autonomous decision, consequence paths, and action commander
- AI debate mode — Growth, Risk, and Finance advisors reconcile to a final recommendation
- Decision scorecard across confidence, urgency, financial impact, operational risk, and complexity
- Industry and use-case framing (Finance, Healthcare, Technology; board, crisis, weekly ops)
- History, share/copy export, executive voice brief with TTS, and onboarding flow
Challenges solved
- Orchestrating five sequential LLM calls with visible progress without losing user trust on latency
- Keeping outputs executive-grade — one primary action, quantified stakes, not generic summaries
- Designing simulated integrations that feel real while clearly scoped as demo-safe
- Shipping a polished mobile + web experience from one Expo codebase for hackathon timelines
Architecture & engineering highlights
Document ingest → full orchestrator (5 Gemini agents) or fast single-call path → structured AnalysisResult in AppContext → Results UI with alerts, debate, scorecard, and simulated execution; history persisted via AsyncStorage.
Engineering highlights
Information → decision
Upload a quarterly report and get a leadership alert, 95% confidence recommendation, and do-nothing vs act-now consequence paths in under a minute.
Transparent agent trace
Step-by-step UI shows each helper’s status, timestamps, and outputs — from reading 144 words to scoring seriousness at 85/100.
Demo-ready execution
Action commander simulates Slack, email, and CRM steps; debate mode surfaces conflicting advisor perspectives before the final pick.
Screenshots
Future improvements
- Live Slack, email, and CRM connectors behind user-approved OAuth
- Team workspaces with shared history and role-based report access
- SHAP-style attribution tying each recommendation to source sentences in the report
- Scheduled re-analysis when new documents arrive in connected drives