Daloy: AI that keeps cities and commuters moving

Problem Statement

Metro Manila is regularly ranked among the most congested cities in the world. JICA has estimated that traffic congestion costs the region billions of pesos per day, approximately PHP 3.5 billion as of its 2017 study, and the figure is projected to grow. Many intersections still run fixed-time signals or are manually overridden by traffic enforcers. Outside Metro Manila, fast-growing cities like Cebu, Davao, and Cagayan de Oro face similar gridlock. Under the PUV Modernization Program, LGUs must prepare Local Public Transport Route Plans (LPTRPs), but many lack the data and analysts to do so, so routes are based on outdated surveys. Commuters wait long hours at terminals while units bunch up or run empty.

Proposed Solution

Daloy has two modules that share one data layer.

1. Signal Optimizer: Computer vision models run on existing LGU/MMDA CCTV streams. They count and classify vehicles (jeepneys, buses, tricycles, motorcycles, cars), estimate queue lengths per approach, and feed a reinforcement learning / optimization engine that recommends green-time splits for each intersection. It runs first in “advisory mode” for traffic engineers, then in automated mode for compatible controllers.
2. Route & Dispatch Planner: It ingests GPS traces from modern PUVs and fleet operators along with passenger demand proxies (ridership counts from CV at terminals, plus OSM and census data). It detects bunching, underserved barangays, and redundant routes, simulates alternatives, and produces LPTRP-ready reports. A dashboard shows congestion heatmaps, signal performance, and before-and-after travel time simulations.

Target Market

• Primary users: MMDA Traffic Engineering Center, city traffic management offices, and LGU planning offices preparing LPTRPs.
• Buyers (B2G/B2B): MMDA, highly urbanized cities (approximately 30+ nationwide), DOTr and LTFRB, transport cooperatives and corporations operating modern PUVs, and private toll and estate developers.
• Size: 17 Metro Manila LGUs plus approximately 140+ cities nationwide, thousands of signalized intersections, and tens of thousands of PUV franchises.

How AI Is Used

• Computer vision (YOLO-class detection plus tracking): Converts raw CCTV into counts, classes, and queue lengths, including Philippine-specific classes such as jeepneys and tricycles. Without it, CCTV is just video that no one watches.
• Reinforcement learning / optimization for signals: Learns timing policies that adapt to real demand. Fixed timing plans cannot react to rush hour, rain, or flooding events.
• Demand forecasting (time series): Predicts flows by hour and weather (e.g., PAGASA rainfall) so the system can pre-adjust signals and dispatch.
• GPS trajectory mining and route optimization: Detects bunching, dwell hotspots, and coverage gaps, and recommends headways and route changes.
• LLM report generation: Summarizes findings into plain-language LPTRP and council briefs for non-technical decision-makers.

PLDT / Smart / ePLDT / Rokid Angle

• PLDT fiber and Smart 5G/LTE: Backhaul for CCTV and signal controllers, plus IoT SIMs for PUV GPS units, giving PLDT Enterprise recurring connectivity revenue.
• ePLDT data centers and edge cloud: GPU inference near the cameras, with sovereign hosting of city video data.
• Smart city B2G sales: Fits PLDT Enterprise’s smart city and safe city portfolio and upsells to existing LGU CCTV deployments.
• Rokid AR (stretch): Traffic enforcers at an intersection see live queue lengths and recommended phase changes on AR glasses.

Demo Scenario (3 minutes for judges)

1. Play a real-looking (public or recorded) video of a congested EDSA-style intersection at 6 PM, with jeepneys and motorcycles everywhere.
2. Daloy overlays detection boxes and live counts per lane: “Northbound queue: 42 vehicles; Eastbound: 6.”
3. A split screen shows a SUMO simulation of the same intersection with fixed timing versus Daloy’s adaptive timing. Average wait drops by about 25–35% in simulation, clearly labeled as simulated.
4. Switch to the Route Planner: GPS traces of a mock modern jeepney route show five units bunched together and a barangay with no service for 40 minutes.
5. Daloy recommends a revised headway and a route extension, and generates a one-page LPTRP brief in English and Filipino.
6. Close: “Every minute saved across a million commuters is thousands of hours given back to Filipino families.”

30-Hour Build Plan

Suggested stack: Ultralytics YOLO (check license, AGPL) or RT-DETR, ByteTrack, OpenCV, Eclipse SUMO with TraCI, Stable-Baselines3 or a max-pressure heuristic, Python/FastAPI, Kepler.gl or Deck.gl, PostGIS, and Gemini for report generation. Public data / APIs: OpenStreetMap (road network), MMDA public traffic information and advisories, LTFRB route and franchise info, DOTr LPTRP manual, PAGASA rainfall, PSA population by barangay, public traffic camera datasets (e.g., UA-DETRAC, plus Philippine dashcam/CCTV clips with appropriate licenses), Hugging Face/Roboflow vehicle datasets that include jeepneys and tricycles, and synthetic GPS traces.

Risks & Privacy Notes

• Sim-to-real gap: Advisory mode first, and pilots at a few intersections with traffic engineer oversight. Simulation results are clearly labeled as simulated.
• Hardware heterogeneity: Many signal controllers are legacy, so the tool outputs recommended timing plans that can be applied manually where automation isn’t possible.
• Data Privacy Act 2012: Only aggregate counts are processed. Plates and faces are not stored and frames are discarded after inference, with edge processing where possible. GPS data is limited to vehicles, not passengers.
• Clause 24 compliance: Only public or licensed videos and synthetic GPS data are used. No identifiable personal data is uploaded to AI tools, and dataset licenses are respected.

AI Use Disclosure (draft)

We use YOLO-family object detection for vehicle counting, a reinforcement learning or optimization controller tested in SUMO simulation, time series forecasting for demand, and an LLM (e.g., Gemini) to draft plain-language reports. AI coding assistants helped scaffold simulation and dashboard code, and some synthetic GPS traces were AI-generated. All outputs were reviewed and validated by the team, and any signal recommendation requires approval by a human traffic engineer. No personal or confidential data was entered into AI tools.

Rubric Scoring