Development Underwriting Engine with CrewAI & NVIDIA NeMo Guardrails
Major developers restarting capital deployment need rigorous risk screening across volatile market and construction signals. Build a CrewAI multi-agent engine wrapped in NVIDIA NeMo Guardrails to synthesize zoning rules, submarket cap rates, and supply-chain risk without hallucinated metrics.
What you are building
The core problem, expected build, and operating context for this challenge.
Orchestrate specialized underwriting roles in CrewAI with NVIDIA NeMo Guardrails safety policies governing financial metric outputs.
How work is evaluated
Evaluates feasibility analysis correctness and checks that NeMo Guardrails block invalid financial metrics.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
guardrail_hallucination_block
Verifies output cap rates are constrained between 2% and 15%
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
memo_accuracy_score
Accuracy of financial ratio calculations against ground truth math models • target: 0.95 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Build a multi-agent crew (Zoning Analyst, Financial Modeler, Risk Manager) using CrewAI.
Define programmable rails in NVIDIA NeMo Guardrails to validate cap rate and debt yield ranges.
Synthesize real estate deal proposals into structured feasibility memos.
Prevent hallucinated land acquisition costs through strict input/output guardrails.
Reference links and supporting material
Synthetic dataset of commercial ground-up development opportunities with site metrics, zoning constraints, and submarket comps.
Challenge setup
Evaluates feasibility analysis correctness and checks that NeMo Guardrails block invalid financial metrics.
Challenge input
JSON containing land_cost, expected_gfa, target_rent_psf, hard_construction_cost
CrewAI
Framework assignment for multi-role agent collaboration.
NVIDIA NeMo Guardrails
Required AI engineering tool for constraining model responses.
Evaluated output
JSON with return_on_cost, dscr_projected, deal_approved boolean
- Verifies output cap rates are constrained between 2% and 15%
- Accuracy of financial ratio calculations against ground truth math models • target: 0.95 • range: 0-1
- Benchmark: Financial-Agent-Safety-Bench
- Memo Accuracy Score target: 0.95
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- RAI · Required
- CrewAI · Optional
- NVIDIA NeMo Guardrails · Optional
Evaluation contract
- guardrail_hallucination_block · Weight 1
- memo_accuracy_score · Weight 1
Metrics
- memo_accuracy_score · Target 0.95
Benchmarks
- Financial-Agent-Safety-Bench
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing land_cost, expected_gfa, target_rent_psf, hard_construction_cost
Output: JSON with return_on_cost, dscr_projected, deal_approved boolean
Execution harness: Python
Evaluation tools
- CrewAI · Multi-Agent Collaboration Framework
- NVIDIA NeMo Guardrails · AI Safety & Output Verification Tool
Recipe freeze marker
No recipe freeze time is recorded.