LifeAudit AI

## Inspiration

Most people have a financial advisor who sees their money, a career coach who sees their job, maybe a therapist who sees their stress. But none of these advisors talk to each other. You can be making perfect decisions in each domain individually while they silently contradict each other — saving aggressively while burning out, upskilling for a career that your financial runway can't support, building a network in an industry you're about to leave.

That gap — the space between your advisors — is where most people get stuck. LifeAudit AI was built to live in that gap.

## What it does

LifeAudit AI runs 5 specialist AI agents in parallel, each independently analyzing a different life domain:

• 💰 Financial Agent — savings rate, runway, debt, hidden waste
• 🚀 Career Agent — market value, skill gaps, trajectory risk
• 🏃 Health Agent — vitality score, burnout risk, highest-ROI interventions
• 🤝 Network Agent — structural gaps, echo chamber risk, missing mentors
• 📚 Learning Agent — future-readiness, critical skills ranked by urgency

After all 5 agents complete, an Orchestrator Agent synthesizes their outputs — not to summarize them, but to find cross-domain conflicts that only emerge when you look across all domains simultaneously.

Example conflict: "Your Career Agent recommends 12hrs/week of upskilling to stay relevant. Your Health Agent independently flagged early burnout signals. Your Financial Agent shows 4 months of runway. These three signals together point to an unsustainable plan — the Orchestrator reorders your action plan to resolve this before anything else."

No single agent could produce this. That's the point.

## How we built it

Built entirely on MeDo, using its multi-agent orchestration capabilities and multi-turn chat system. The architecture was designed so agents run with isolated context — they cannot see each other's reasoning, only their own domain inputs. This is intentional: cross-contamination of reasoning degrades each specialist's analysis quality.

The Orchestrator only receives agent outputs, never their internal chains. It then runs a conflict-detection pass across all 5 outputs before generating the Life Score and action plan.

The "What If" simulation feature re-triggers the relevant specialist agent with updated inputs, then passes the new output back to the Orchestrator to check whether conflicts are resolved or new ones emerge — demonstrating that the multi-agent system is doing real dynamic work, not just rendering static results.

## Challenges

The hardest architectural challenge was ensuring agents didn't produce generic advice. A single prompt asking "analyze my finances and career" produces vague, hedged output. Isolating each agent with a strict domain boundary and then synthesizing at the orchestration layer produced dramatically sharper, more specific analysis.

The second challenge was making the multi-agent architecture visible to users — not just a black box that returns a score. The reasoning log on each agent card during the analysis screen was the solution: users can watch each agent think through their domain in real time, then see the Orchestrator connect the dots.

## What we learned

Multi-agent systems are only as powerful as their orchestration layer. Parallel agents without a strong synthesis step are just parallel prompts. The value is entirely in what the Orchestrator does with the conflict space between agent outputs — that's where the insight lives that no single AI, and no single human advisor, can produce alone.

Built With

• deno
• eslint
• framer-motion
• google-ai-gateway
• google-gemini-2.5-flash
• montserrat
• pnpm
• postgresql
• prettier
• react-19
• react-router
• shadcn/ui
• supabase
• supabase-edge-functions
• supabase-row-level-security
• tailwind-css
• typescript
• vite