Sentinel

• 

  main dashboard

• 

  document ingestion screen

• 

  document reviewed screen

Inspiration

CRE acquisition analysts spend 4–8 hours screening every broker deal, and most die at LOI anyway — because broker claims go unchecked and the data needed to verify them is scattered across 10+ public sources nobody can cross-check in time.

What it does

Sentinel takes a broker's offering memorandum and autonomously investigates it against public records — owner history, loans, permits, tax status, code violations, and comps — then returns a defensible pursue/watchlist/pass verdict in under 90 seconds, reasoning out loud the whole way.

How we built it

Built on Hermes Agent driving NVIDIA Nemotron 3 (via NVIDIA NIM). Every data source is a uniform, self-describing JSON job the agent discovers and calls itself, so the orchestration is the model's own reasoning rather than a hardcoded script.

Challenges we ran into

Designing the agent to investigate rather than just summarize, and keeping the demo deterministic under free-tier API rate limits.

Accomplishments that we're proud of

connecting agents to the world

What we learned

Agent-native architecture — uniform JSON job contracts — makes autonomous tool use dramatically more reliable than glued-together code.

What's next for Sentinel

getting this in the hands of real syndicates

Built With

• hermes-agent
• nvidia-nemotron
• nvidia-nim
• pdfplumber
• pypdf
• python
• uv