Cashseer

Dashboard
Insights ( Dept)
Negotiator
Grants
Risk Analyser

Inspiration

47 million small businesses in the US operate without a CFO. 82% of business failures are caused by poor cash flow management. A real CFO costs $300,000 per year — completely out of reach for small business owners. We built Ghost CFO to change that forever.

What it does

Ghost CFO is an autonomous multi-agent AI system that acts as a full-time CFO for small businesses — not by giving advice, but by taking real financial actions on their behalf.

It does 4 things automatically:

Forecasts cash crises 45 days early with week-by-week predictions
Detects financial danger with a Crisis Score from 0-100
Hunts government grants by literally navigating grants.gov live using Amazon Nova Act browser automation
Negotiates with suppliers by drafting and sending real emails autonomously via Gmail

When the Crisis Score exceeds 70, Ghost CFO activates all recovery agents automatically — no human input required.

How we built it

Ghost CFO is built on a multi-agent architecture powered entirely by Amazon Nova 2 Lite and Amazon Nova Act:

Orchestrator Agent (Nova 2 Lite) — reads the financial situation and routes tasks to specialist agents
Cash Flow Agent (Nova 2 Lite) — ingests CSV transaction history, generates 45-day forecast
Crisis Predictor(Nova 2 Lite) — scores danger 0-100, triggers recovery automatically
Grant Hunter (Nova Act) — navigates grants.gov live in a real browser, finds matching grants
Supplier Negotiator (Nova Act + Nova 2 Lite) — drafts and sends real negotiation emails via Gmail

Backend: FastAPI on AWS EC2. Frontend: Streamlit dashboard with Plotly charts. All agents communicate through structured JSON.

Challenges we ran into

Amazon Nova 2 Lite requires cross-region inference profiles — the model ID fix from amazon.nova-2-lite-v1:0 to us.amazon.nova-lite-v1:0 was our first major breakthrough
Nova Act browser sessions are non-deterministic — government websites change layouts frequently, requiring robust fallback logic
Wiring 5 agents together without race conditions required careful pipeline architecture and retry logic throughout
Getting Nova 2 Lite to return consistently valid JSON required extensive prompt engineering ## Accomplishments that we're proud of
Built a system that takes REAL actions — not just generates advice
Nova Act navigating a live government website autonomously is something no other team attempted
The autonomous email sending via Gmail with zero human input
Crisis Score triggering automatic recovery across all agents simultaneously
A dashboard that looks and feels like a real financial product ## What we learned
Amazon Nova Act is genuinely revolutionary — browser automation at this level changes what AI can do in the real world
Multi-agent orchestration is harder than single-model solutions but infinitely more powerful
Prompt engineering is as important as Python logic — the system prompt IS the agent's brain
RAG (injecting real CSV data into prompts) makes the model reason over actual business data, not hallucinations