6. Runtime View

Scenario 1: CRESI Calculation Flow

User Goal: Lender submits a valuation request and receives a sentiment-adjusted appraisal.

Sequence Diagram

Step-by-Step Breakdown

Step	Component	Action	Duration	Error Handling
1	WebApp	User fills form, submits	<1s	Client-side validation
2	APIGateway	JWT validation, rate limit check	<50ms	401 Unauthorized or 429 Too Many Requests
3	Smitty	Retrieve historical weights from Memory Bank	<100ms	Fallback to default weights
4	CRESI	Parallel fetch 6 layer data sources	<5s	Timeout 30s, fallback to cached data
5	CRESI	Normalize and weight layer scores	<100ms	N/A (deterministic calculation)
6	Smitty	Generate human-readable explanation	<500ms	Return raw scores if explanation fails
7	APIGateway	Async store valuation in Memory Bank	<200ms	Retry queue if DB unavailable
8	WebApp	Render results dashboard	<500ms	Retry button if render fails

Total Latency: 6-10 seconds (target: <10s for 95th percentile)

---

Scenario 2: Weight Optimization (Nightly Batch)

Goal: Smitty learns from previous day's valuation outcomes to improve future predictions.

Sequence Diagram

Weight Update Logic

# Pseudocode for weight optimization
def optimize_weights(outcomes, current_weights, learning_rate=0.01):
    mape = calculate_mape(outcomes)

    for layer in ['macro', 'micro', 'geo', 'capital', 'demo', 'events']:
        gradient = calculate_gradient(layer, outcomes)
        new_weight = current_weights[layer] - learning_rate * gradient

        # Constraints
        new_weight = max(0.05, new_weight)  # Minimum 5% weight

    # Normalize to sum to 1.0
    weights = normalize_weights(new_weights)

    return weights

Outcome Data Requirements:

• Actual market outcome (sale price, lease rate, occupancy)
• Time between valuation and outcome (typically 3-12 months)
• Prediction confidence at time of valuation

MAPE Alert Thresholds:

• Green: MAPE < 5% (excellent)
• Yellow: MAPE 5-10% (acceptable)
• Red: MAPE > 10% (investigate layer data sources)

---

Scenario 3: Layer Data Fetch Timeout Handling

Problem: External API (e.g., CoStar) times out during CRESI calculation.

Sequence Diagram

Fallback Strategy:

Data Freshness	Confidence Impact	Action
< 1 day	No impact (100%)	Use cached data as-is
1-7 days	-10% confidence	Apply staleness discount
7-30 days	-25% confidence	Warn user of stale data
> 30 days	Fail layer	Exclude layer from calculation, notify admin

Circuit Breaker:

• After 3 consecutive timeouts: Mark API as degraded
• Switch to cache-only mode for 15 minutes
• Send PagerDuty alert to on-call engineer
• Auto-recovery after successful API call

---

Scenario 4: Explainable AI Report Generation

Goal: Generate regulatory-compliant explanation of how sentiment score was calculated.

Data Flow

Explanation Components

1. Layer Contribution Table:

Layer                Score    Weight    Contribution    Importance
───────────────────────────────────────────────────────────────────
Macro-Economic       +15      0.25      +3.75          High
Micro-Economic       -5       0.30      -1.50          High
Geopolitical         +2       0.10      +0.20          Low
Capital Markets      +20      0.20      +4.00          High
Demographics         +8       0.10      +0.80          Medium
Current Events       +5       0.05      +0.25          Low
───────────────────────────────────────────────────────────────────
FINAL SCORE: +7.50 (Bullish sentiment)

2. Plain English Narrative:

"The CREstimate.ai sentiment score for this Chicago office property is +7.5 (slightly bullish). This is driven primarily by strong capital markets conditions (investor appetite for office assets remains high) and positive macro-economic trends (GDP growth, declining interest rates). However, local micro-economic factors are negative due to rising vacancy rates in the Chicago CBD submarket. The model confidence is 82%, indicating high reliability of this prediction."

3. SHAP Values Visualization:

• Waterfall chart showing each layer's push/pull on final score
• Feature importance ranking
• Comparison to historical average for this asset type/location

4. Regulatory Compliance Metadata:

• Data sources cited (with timestamps)
• Model version and training date
• Weight optimization history
• Confidence interval calculation method

---

Performance Characteristics

Latency Targets

Operation	Target	P95	P99	Notes
Full valuation	<10s	8s	12s	Includes all 6 layers
Single layer fetch	<2s	1.5s	3s	Individual API call
Weight optimization	<5min	3min	10min	Nightly batch
Explanation generation	<500ms	400ms	1s	Post-calculation
Cache lookup	<50ms	30ms	100ms	Redis in-memory

Throughput Targets

Metric	Target	Notes
Concurrent users	100	Per instance
Valuations/hour	500	Peak load
API requests/sec	50	Rate limit per client
Cache hit rate	>80%	For layer data

Failure Modes

Failure	Detection	Recovery	SLA Impact
Single layer timeout	30s timeout	Use cached data	No impact if cache hit
All layers timeout	3 consecutive failures	Circuit breaker, alert	Degraded service
Memory Bank unavailable	Health check failure	Queue valuations, retry	No immediate impact (async)
Weight optimizer crash	Cron job failure	Use previous weights	No immediate impact (daily batch)

---

For component details, see Building Block View. For deployment infrastructure, see Deployment View.