# Demo Gallery

> Visual walkthrough of the NYC Taxi Zone Recommendation platform.

---

## Scenario: Rainy Friday Evening in Manhattan

### The problem

A taxi driver finishes a drop-off in Midtown at 7:30 PM on a rainy Friday. Without guidance, the driver cruises randomly, wasting fuel and time.

**Before AI guidance:**
- Driver circles Midtown for 20 minutes
- Finds passenger heading to Brooklyn ($18 fare)
- 35% utilization across the shift
- ~$350 daily revenue

### AI decision process

```
Step 1: Demand forecast → JFK airport demand spikes at 8 PM (rain + Friday)
Step 2: Travel time matrix → 35 min from Midtown to JFK via highway
Step 3: Two-step horizon planner → Go to JFK now, pick up airport fare, reposition to Manhattan
Step 4: Recommendation → [JFK Zone 132, Upper East Side Zone 140, Midtown Zone 161]
```

### The result

**After AI guidance:**
- Driver heads to JFK, picks up $62 airport fare within 10 minutes
- Then repositions based on next forecast window
- 52% utilization across the shift
- ~$570 daily revenue (+$220/day vs cruising)

---

## Benchmark Dashboard

### Static diagnostic (3,360 queries)

![NDCG comparison](../assets/ndcg_comparison.png)

The Two-Step Horizon strategy achieves 0.9565 NDCG@3 on public validation queries — a 21.9% improvement over the naive Hot Zone baseline.

### Rollout performance

![Pickup comparison](../assets/pickup_comparison.png)

100-seed paired rollout shows consistent improvement: Two-Step delivers +$139/day vs Hot Zone, with the improvement concentrated in the 7-10 AM and 6-9 PM peak windows.

### Multi-agent competition

The multi-agent simulator reveals that as fleet size grows, simpler strategies degrade faster than horizon-aware policies. At 50 drivers, Two-Step maintains a +$25/day advantage over Single-Step.

---

## NYC Map Visualization

The project uses NYC's 263-taxi-zone geography. All-pairs travel times are precomputed via Dijkstra on a directed OD graph built from 1.8M+ training trips.

```
        Manhattan (Zones 100-199)
        ┌──────────────────────────┐
        │  Upper East   Upper West │
        │  ┌────┬────┐ ┌────┬────┐ │
        │  │140 │141 │ │142 │143 │ │
        │  └────┴────┘ └────┴────┘ │
        │  Midtown                 │
        │  ┌────┬────┐             │
        │  │161 │162 │   → Queens  │
        │  └────┴────┘      (JFK)  │
        │  Downtown          ↓     │
        │  ┌────┬────┐     ┌───┐   │
        │  │113 │114 │     │132│   │
        │  └────┴────┘     └───┘   │
        └──────────────────────────┘
```

---

## Reproducibility

All results in this gallery are reproducible. Run:

```bash
make all          # Full pipeline
make static       # Static diagnostic only
make combined-benchmark  # Combined report
```

All metrics are checked in as reference snapshots in `outputs/` with timestamped validation.

---

## More scenarios

| Scenario | Time | Weather | AI Decision | Outcome |
|---|---|---|---|---|
| Monday morning rush | 8:15 AM | Clear | Upper East → Midtown | Commuter demand |
| Saturday night | 11:00 PM | Clear | Greenwich Village → Meatpacking | Nightlife flow |
| Airport surge | 4:00 PM | Thunderstorm | JFK → Manhattan | Airport backlog |
| Holiday eve | 6:00 PM | Snow | Penn Station → residential | Transit hub exit |