Our Mission

To harness artificial intelligence to improve waste management and recycling systems globally. By leveraging intelligent technologies, the project reduces waste, increases recycling efficiency, optimises resource use and minimises environmental impact across cities, industries and communities.

Our Vision

To become the leading AI-based solution in intelligent waste management — enabling smarter cities, cleaner environments and a truly circular economy. We envision a world where waste is minimised, resources are reused efficiently, and technology actively protects the planet for future generations.

How the System Works

The platform collects and analyses data from waste streams, recycling facilities and smart sensors in real time. AI algorithms identify patterns, optimise waste sorting, improve recycling efficiency and support smarter decision-making — automating analysis so cities can reduce waste, lower costs and minimise environmental impact without additional human overhead.

Market data: DataRoot Labs, 2025 Waste Management Technology Landscape; Columbia Climate School, June 2025.

Key Industry Milestones (2024–2025)

• Greyparrot's AI analytics detected over 40 billion waste objects across 180+ systems in 20 countries in 2024 — on pace for 100 billion by end of 2025.
• AMP Robotics deployed 400+ AI systems worldwide, raised $91M; their robots identify 50+ material categories using deep learning trained on millions of images.
• Antfarm X1 (Amsterdam) sorts 700 items per minute. AMP ONE processes thousands per minute via air-jet sorting.
• Bollegraaf — the world's largest recycling facility builder — is retrofitting thousands of plants with AI in partnership with Greyparrot.
• AI adoption in waste management is growing at 22% per year through 2030.

Economic Case for Cities

Smart sensors eliminate unnecessary collection trips, reducing fuel consumption and operational costs by up to 30% in municipalities that have deployed them. The broader circular economy approach could generate up to $4.5 trillion in annual economic output by 2030 globally.

Regulatory Drivers

• EU Waste Framework Directive (2018/851) targets 65% recycling of municipal waste by 2035.
• Extended Producer Responsibility (EPR) laws expanding across Europe hold brands financially accountable for end-of-life packaging.
• EU landfill penalties and the Carbon Border Adjustment Mechanism make inaction increasingly costly for municipalities and industries.

Sources: DataRoot Labs (2025); Columbia Climate School (2025); Resource Recycling (Dec 2025); Okon Recycling (Aug 2025).

Bin-Level Sensors

• Ultrasonic / LiDAR fill-level sensors — continuous monitoring, alerting collection services only when bins approach capacity. Eliminates needless collection trips — proven to cut operational costs up to 30%.
• Embedded NIR Spectrometers — near-infrared composition analysis identifies waste type (organic, plastic, metal, glass, e-waste) before collection, enabling pre-sorted routing to the right facilities.
• RFID Identification — links each bin deposit to the citizen incentive blockchain system, enabling per-household tracking without privacy exposure.
• Gas Analysers (CH₄, CO₂) — detect fermentation and hazardous build-up in organic waste bins before overflow or odour incidents occur.

Vehicle-Level Sensors

• Real-time GPS for live fleet position and route compliance monitoring.
• Compaction sensors measuring load density per trip to optimise vehicle fill efficiency.
• On-board gas analysers tracking per-vehicle emissions for carbon accounting.

Connectivity Stack

• LoRaWAN / NB-IoT (LPWAN) — ultra-low power, wide-area coverage for all fixed bin sensors. Years of battery life per unit. One gateway covers hundreds of sensors.
• 5G — for real-time video from sorting cameras and high-bandwidth vehicle telemetry requiring sub-millisecond latency.

Source: Okon Recycling (2025); MDPI Sensors (2025); API4AI (2025).

LSTM Waste Volume Forecasting

Long Short-Term Memory (LSTM) time-series models analyse historical fill rates, weather data, local event calendars and seasonal patterns to forecast waste peaks. Inputs include: daily/weekly demand cycles, neighbourhood demographics, upcoming events (markets, festivals, public holidays) and weather forecasts (heat accelerates organic waste fermentation).

Q-Learning Route Optimisation

Reinforcement learning agents treat route planning as a sequential decision problem, continuously learning from real-world outcomes. The agent simultaneously optimises:

• Fill-level urgency (prevents overflow incidents)
• Fuel cost and CO₂ minimisation per vehicle
• Traffic condition integration via live APIs
• Global commodity prices to prioritise high-value material stream collection

Predictive Maintenance

IoT sensors on vehicles and sorting equipment monitor temperature, vibration, pressure and load. AI-driven forecasting detects anomaly signatures weeks before equipment failure — preventing costly unplanned breakdowns and service disruptions.

Sources: SWANA (2023); API4AI Medium (2025); MDPI SMILE (2025).

How It Works

• Each household receives a unique QR/barcode linked to their registered city platform account.
• When recyclables are deposited at smart bins, NIR sensors and cameras verify correct sorting in real time.
• Each verified action is recorded as an immutable transaction on Hyperledger Fabric blockchain.
• Verified points convert directly to property-tax reductions of up to €85/year — transparent, auditable, tamper-proof.

Bonus / Malus System

Households that consistently contaminate recycling streams receive automated bin-side alerts, followed by graduated point reductions. This creates a city-wide norm of responsible sorting through positive economic incentives — not punitive enforcement.

Why Behavioural Design Matters

Public confusion about recyclability is a primary driver of contamination rates globally. Our platform addresses this through:

• Real-time bin-side visual feedback on sorting accuracy at the moment of deposit
• A companion mobile app showing household environmental impact, points balance and neighbourhood leaderboards
• Gamified community challenges backed by behavioural science research on habit formation and social norms

Industry Benchmarks (2025)

• Antfarm X1 (Amsterdam) — 700 items per minute, fully automated.
• AMP ONE (Cleveland) — identifies 50+ material categories via air-jet sorting, processing thousands of items per minute.
• Greyparrot — 111 distinct waste categories; detected 40+ billion objects in 2024 across 180 systems in 20 countries; 100 billion+ projected for 2025.
• Glacier (formerly EndWaste) — compact AI robotic sorters for decentralised small-scale MRFs, making recycling economically viable for smaller municipalities.

Our Sorting Stack

• Edge AI (MobileNetV3) — embedded in bin cameras for instant pre-classification at the point of deposit, before material reaches the sorting centre.
• Cloud AI (ResNet-152 + Vision Transformer) — deep networks trained on 500,000+ annotated waste images across 50+ categories. The ViT component captures long-range spatial relationships that CNNs miss — critical for mixed-composition items.
• Hyperspectral cameras — identify polymer types within mixed plastics and detect material composition beyond visible light at sorting centres.
• AI robotic arms — recognise materials by colour, texture, shape and brand label, sorting plastics, metals, glass, organics and electronics at high throughput.

Self-Improving Architecture

GANs (Generative Adversarial Networks) generate synthetic training data for rare waste anomalies. The model becomes more accurate with each week of operation — adapting automatically to new packaging materials as they enter the waste stream.

Sources: Columbia Climate School (2025); Resource Recycling (Dec 2025); PYMNTS (2024); DataRoot Labs (2025).

Material Chain of Custody

Every kilogram of material — from bin deposit → sorting centre → recycling plant → new product — is recorded as a tamper-proof on-chain transaction. This creates:

• Automated EPR (Extended Producer Responsibility) compliance reporting for brands
• Real-time contamination source identification (which bin, which area, which material stream)
• Brand-level packaging accountability: AI can identify brand logos on items, enabling producer responsibility enforcement
• Immutable audit trail for city regulators and EU environmental certification bodies

Verified Carbon Credits (MRV Model)

Our Monitoring, Reporting, Verification model is aligned with IPCC standards. For each tonne of CO₂-equivalent avoided: sensors provide monitoring data, AI models calculate verified avoidance, blockchain provides the unalterable verification record. Credits can be sold on voluntary carbon markets or applied to city carbon neutrality targets.

Digital Twin Simulation

Monte Carlo simulations run against our digital twin of Toulouse's waste system model circular economy scenarios — for example, assessing the impact of new bioplastics mandates or EPR fee changes before they take effect, enabling proactive policy design.

Pipeline Architecture

• Apache Kafka — ingests high-frequency streams from 10,000 bins and the vehicle fleet simultaneously, with guaranteed delivery and full event replay capability.
• Spark Streaming — processes events in milliseconds, triggering real-time alerts and ML model inference. Detects overflow risk, contamination spikes and route deviations instantly.
• Feature Store (Feast) — maintains consistent feature engineering across training and serving pipelines, eliminating model drift — a leading cause of AI production failures.
• TensorFlow Serving — hosts all production models with versioning, A/B testing capability and sub-10ms inference latency.

Key Performance Indicators

• Purity Index — measures contamination rate per waste stream using ML-based clustering of sorting error patterns. Identifies problem sources down to street level.
• Carbon Avoidance Score — real-time CO₂ accounting per collection cycle, IPCC MRV-aligned, exportable for carbon credit certification.
• Citizen Engagement Metrics — per-household sorting compliance, points balance trends and neighbourhood leaderboards updated in real time.

Open API Interoperability

All data is accessible via standardised Open APIs compatible with SAP EWM, existing smart city platforms and Toulouse's urban planning tools — ensuring our system enhances, not replaces, the city's existing digital infrastructure.

Environmental Impact

• 500 tonnes of avoidable CO₂ eliminated from the collection fleet in Year 1 via route optimisation alone.
• Circular economy material recovery replaces virgin resource extraction — with measurable reductions in upstream industrial emissions for each tonne recovered.
• Digital twin simulations enable future policy decisions (bioplastics mandates, EPR changes) to be stress-tested before implementation.

Economic Impact

• €25M annual waste budget — projected 30% reduction in operational costs = €7.5M in annual savings.
• Carbon credit revenues from IPCC-verified MRV model (estimated €50/tonne CO₂ on voluntary markets).
• Avoided EU landfill penalty exposure worth millions annually as 2035 EU targets approach.
• Full ROI within 3 years of deployment completion.

Social Impact

• Up to €85/year returned directly to participating households — most impactful for lower-income residents.
• Improved air and ground quality through reduced landfill gas and incineration particulates in city districts.
• New green-tech jobs in sensor maintenance, AI model operations, data analysis and citizen engagement.

Global Benchmarks for Context

Amsterdam targets a 50% reduction in new raw material use by 2030. Cities with smart sensor deployments achieve up to 30% operational cost reductions. Our model builds directly on these proven results, adapted for Toulouse's 500,000+ population scale.

Sources: Okon Recycling (2025); MDPI SMILE Conference (2025); Columbia Climate School (2025); SWANA (2023).

Bin Level (10,000 Units)

• Ultrasonic / LiDAR fill sensors — continuous monitoring, no moving parts, 5+ year battery life on LoRaWAN networks.
• Embedded NIR spectrometers — identifies plastics, organics, metals and glass by near-infrared reflectance without sample extraction. Latest 2025 miniaturised NIR chips achieve sorting-centre accuracy at bin size and cost.
• RFID readers — link bin deposits to household accounts for incentive tracking without revealing personal data.
• CH₄ / CO₂ gas analysers — fermentation and hazard detection in organic bins; triggers priority alerts before overflow or odour incidents.

Sorting Centre Level

• Intelligent weighing systems with per-stream material granularity.
• Hyperspectral cameras (extending beyond NIR into SWIR) for polymer-type differentiation within mixed plastics — distinguishing PET from HDPE from PVC at conveyor speed.
• Computer vision arrays feeding real-time classification to edge AI models on-site.

LPWAN — LoRaWAN / NB-IoT

Low-Power Wide-Area Networks for all fixed bin sensors: ultra-low power consumption (years per battery charge), 15km urban range, deep building penetration. One LoRaWAN gateway covers hundreds of sensors at minimal recurring cost. Proven at scale in hundreds of European smart city deployments.

5G — High-Bandwidth Applications

Real-time video from sorting centre cameras and on-vehicle telemetry require 5G's multi-Gbps throughput and sub-1ms latency. 5G also enables on-vehicle edge AI inference — processing camera feeds without cloud round-trip delays for time-critical sorting decisions.

Network Resilience & Federated Updates

Multi-path routing ensures fill-level alerts are delivered even during network congestion events. Federated learning propagates AI model improvements across the sensor network without centralising raw data — a key privacy and efficiency advantage over cloud-only architectures.

Why Edge AI for Waste?

Sending raw video from 10,000 bins to the cloud for every deposit is bandwidth-intensive, slow and privacy-sensitive. Edge AI runs compact neural networks directly on the bin's embedded processor — delivering instant classification with no network round-trip and no raw video leaving the device.

MobileNetV3 Architecture

• Squeeze-and-Excitation (SE) blocks — channel-wise attention that boosts accuracy without parameter overhead.
• Hard-Swish activation — computationally cheaper than standard Swish; optimised for ARM Cortex hardware acceleration.
• Neural Architecture Search (NAS) — automatically discovers the optimal architecture for target latency budgets on embedded processors.
• INT8 Quantization-Aware Training — reduces model size 4× with less than 1% accuracy degradation. Fits in 4MB of flash storage.

2025 Research Validation

Research published in MDPI Sensors (June 2025) confirms that quantized CNN models deployed on IoT devices achieve high-accuracy waste classification at the point of disposal — in smart bins, mobile apps and household assistive tools. Lightweight models continuously improve as new waste materials enter the stream via over-the-air updates.

Source: MDPI Sensors, "AI for Sustainable Recycling: Efficient Model Optimization for Waste Classification Systems" (June 18, 2025).

ResNet-152

152-layer residual network trained on our 500,000+ image proprietary waste dataset across 50+ material categories. Residual skip connections solve the vanishing gradient problem, enabling very deep training to distinguish visually similar materials — such as PET vs HDPE plastics or aluminium vs stainless steel.

Vision Transformers (ViT) — 2024/2025 State of the Art

ViTs apply transformer self-attention to image patches, capturing long-range spatial dependencies that convolutional networks miss. For waste classification, this is critical when identifying materials from partial views or overlapping compositions on high-speed conveyor belts. Our hybrid ResNet-ViT ensemble achieves best-in-class accuracy on benchmark waste classification datasets.

Transfer Learning for Rare Cases

Industrial composites, medical sub-waste types and novel packaging materials are rare in training data. Transfer learning from large-scale ImageNet pre-training, combined with GAN-generated synthetic rare-class samples, ensures the model handles infrequent edge cases as reliably as common materials.

Industry Scale Reference

AMP's deep learning system — the same architectural approach — identifies 50+ categories and continuously trains on millions of images across 400+ deployed facilities. Greyparrot's cloud AI classifies 111 waste categories in real time across 180+ recycling plants globally.

Sources: Columbia Climate School (2025); Resource Recycling (Dec 2025); PYMNTS (2024).

Q-Learning Reinforcement Agent

The agent treats route planning as a Markov Decision Process: at each timestep, it selects which bins to visit next to maximise a reward function that simultaneously balances:

• Fill-level urgency scores (prevents overflow at critical bins)
• Fuel cost and per-vehicle CO₂ emission budgets
• Real-time traffic conditions via live city APIs
• Current global commodity prices (prioritises high-value material stream collection when prices are favourable)

Multi-objective linear programming enforces hard constraints (vehicle capacity, driver hours, emission caps) while the RL agent optimises across soft objectives.

LSTM Demand Forecasting

LSTM networks process time-series fill-rate data from each of the 10,000 bins, incorporating seasonal cycles, local event calendars, weather forecasts and neighbourhood demand patterns. Output: per-bin fill probability curves for the next 72 hours — enabling proactive rather than reactive fleet dispatch.

Fleet Orchestration

A dynamic bin-packing algorithm for vehicle routing ensures each truck operates near full-load efficiency — reducing total vehicle-kilometres by 20% versus fixed schedules while maintaining zero overflow events.

Pipeline Stages

• Apache Kafka — distributed event streaming that ingests data from 10,000 sensors, vehicle GPS and sorting cameras simultaneously. Handles millions of events/second with guaranteed delivery and full replayable event history for model retraining.
• Apache Spark Streaming — distributed real-time processing aggregates, transforms and enriches raw events into structured features within milliseconds. Fires immediate alerts for overflow risk, sensor anomalies or route deviations.
• Feature Store (Feast) — central repository of engineered features shared between offline model training and online production serving. Eliminates training-serving skew — the most common source of AI production failures.
• TensorFlow Serving — production model server hosting all classification and optimisation models with versioning, A/B testing and consistent sub-10ms inference latency under load.

Advanced KPIs Generated in Real Time

• Purity Index — per-stream contamination rate using ML clustering of sorting error patterns, updated continuously.
• Carbon Avoidance Score — IPCC-aligned CO₂ equivalents avoided per collection cycle, exportable for carbon credit certification.
• Fleet Efficiency Score — actual vs optimal routes, load factors and emission performance per vehicle.

Why Hyperledger Fabric?

• Permissioned network — only authorised participants (city authority, recycling centres, certification bodies) can write records, maintaining data integrity without public exposure.
• Pluggable Byzantine Fault Tolerant consensus — adapted for low-latency city-scale transaction volumes (thousands of bin deposits per minute).
• Smart contracts (chaincode) — automate reward distribution, EPR compliance logging and carbon credit issuance without manual intervention.
• Channel architecture — selective data sharing: recycling centre composition data visible to regulators; household deposit data visible only to the city tax authority.

Verified Carbon Credits via MRV

Our Monitoring, Reporting, Verification model aligned to IPCC standards creates a three-layer verification chain: sensors provide raw monitoring data → AI models calculate avoidance quantities → blockchain provides the immutable, timestamped verification record. Resulting credits are tradeable on voluntary carbon markets.

Producer Accountability

Because AI can identify brand logos on packaging items, the blockchain can record which brands' packaging appears most frequently in contaminated waste streams — enabling regulators to enforce EPR obligations based on verified data rather than producer self-reporting.

Homomorphic Encryption

Homomorphic encryption allows computation on encrypted data without ever decrypting it. Our AI models analyse sensitive medical or industrial waste data — producing actionable insights — without any party seeing raw private information. This breakthrough enables GDPR-compliant AI analytics even on the most sensitive data categories (clinical waste from CHU facilities, hazardous industrial byproducts).

Federated Learning

Instead of centralising all sensor and facility data, federated learning trains AI models locally at each site — then aggregates only the encrypted model weight updates, never the raw data. Benefits:

• Industrial partners retain full proprietary control over their waste composition data
• Models improve from distributed real-world data without data leaving its origin site
• Reduces central data storage costs and GDPR compliance exposure
• Enables collaboration between competing facilities on shared model improvements without data sharing

Citizen Data Protection

Household-level data is pseudonymised at source using a barcode-to-token mapping architecture. Only the city tax authority can resolve tokens to household identities for tax reduction processing — no third party, including our own platform, can link waste behaviour data to individual citizens without explicit, auditable consent.

Role in the Project

• Authorisation and coordination for sensor deployment across 10,000 public and semi-public bin locations citywide.
• Integration with Toulouse's GIS, traffic management and urban planning platforms via Open API, connecting Smart Waste AI into the broader smart city ecosystem.
• Administration of the household tax-reduction incentive mechanism — bridging blockchain-verified sorting records to property tax adjustments through the city finance authority.
• Policy leadership toward Toulouse's 2030 zero-waste target, positioning the city for EU Horizon Europe and cohesion fund financing eligibility.

Strategic Alignment

This project directly supports Toulouse Métropole's environmental commitments and aligns with the EU Circular Economy Action Plan, the European Green Deal targets and France's national waste reduction roadmap (loi AGEC). A successful Toulouse pilot creates a replicable template for every mid-size European city facing the same challenges.

INSA Toulouse Contributions

• Development and validation of LSTM waste forecasting models and Q-Learning route optimisation agents — with published methodology making Toulouse a global reference site.
• Dataset curation: materials science expertise guides annotation of the 500,000+ waste image training corpus, ensuring classification categories align with real industrial recycling processes.
• Edge AI optimisation: adapting ResNet and MobileNetV3 architectures for deployment on the specific embedded processors used in Toulouse's sensor hardware.
• Continuous model monitoring and retraining as new materials enter the waste stream — maintaining accuracy over a multi-year deployment lifetime.

CHU Toulouse Contributions

• Medical waste classification protocols: ensuring AI correctly handles all hazardous healthcare waste sub-categories with full ANSM and European regulatory compliance.
• Epidemiological impact assessment: quantifying the public health benefits of reduced landfill gas and incineration particulates in Toulouse districts adjacent to current waste infrastructure.
• Privacy and ethics framework for sensitive waste data, including homomorphic encryption implementation guidance for clinical waste streams.

The Citizen Role

In traditional waste systems, citizens are the last mile of quality control — and the least supported. Public confusion about recyclability drives contamination rates globally. Our platform inverts this dynamic completely:

• Real-time bin-side visual and audio feedback tells residents whether their deposit was correctly sorted — immediately, at the moment it matters most.
• The companion app shows each household's cumulative environmental impact (CO₂ saved, materials recovered), earnings history and neighbourhood leaderboard position.
• Gamified community challenges create positive social norms — backed by peer-reviewed behavioural science on habit formation, social proof and commitment mechanisms.

Economic Equity

Up to €85/year in property tax reductions is meaningful income — particularly for lower-income households. This makes the system socially equitable: those who engage most with the platform benefit most financially, regardless of wealth. The project was designed with this equity principle as a non-negotiable constraint.

Co-Design Process

Residents from the pilot districts participated directly in designing the incentive system, app interface and bin-side communication — ensuring the solution fits real human behaviours and not just theoretical behavioural models. This participatory approach is itself a best-practice signal for EU smart city funding evaluation criteria.

Capital Allocation

• IoT Hardware (40% — €480k) — 10,000 smart bin sensors (ultrasonic, NIR, RFID, gas), LoRaWAN gateway network, vehicle telemetry units, installation and commissioning.
• AI Platform & Cloud (30% — €360k) — GPU compute for model training, TensorFlow Serving deployment, Kafka/Spark cluster, Feast Feature Store, Power BI licensing.
• Blockchain & Incentive System (15% — €180k) — Hyperledger Fabric network deployment, smart contract development, citizen app build, city tax authority integration.
• Integration, Training & Contingency (15% — €180k) — SAP EWM and smart city API connections, operator training programme, 12-month model maintenance reserve.

Revenue & Savings Breakdown

• Collection operational savings: €7.5M/year (30% of €25M current budget) from route optimisation and eliminated unnecessary trips.
• EU landfill penalty avoidance: increasing annually as EU 2035 recycling mandates tighten — currently €100+/tonne in penalty exposure.
• Carbon credit revenues: MRV-verified credits at ~€50/tonne CO₂ on voluntary markets. 500 tonnes CO₂ avoided in Year 1 = €25k initial credit revenue, scaling with system expansion.
• EU grant eligibility: full deployment unlocks access to Horizon Europe smart city funding and EU cohesion fund financing — potentially covering future expansion phases.

Technology Risk Profile

All core components are commercially proven at scale: LoRaWAN sensor networks (deployed in hundreds of European cities), ResNet/LSTM AI (powering 400+ AMP facilities globally), Hyperledger Fabric (enterprise standard across finance, supply chain and public sector). Primary integration risk is mitigated by our phased deployment approach — pilot district first, full city rollout validated before full investment commitment.

Financial model based on: Okon Recycling (2025); AMP Robotics deployment data (2024); SWANA municipal benchmarks (2023).

Vision & Founding Motivation

Adam launched Smart Waste & Recycling AI with a conviction that the environmental crisis cannot be solved by awareness campaigns alone. It requires systemic, technology-driven transformation of the infrastructure through which cities interact with their waste — building solutions that are economically self-sustaining and scalable, not dependent on perpetual subsidy or goodwill.

His approach deliberately bridges academic rigour (partnership with CHU and INSA Toulouse) with entrepreneurial urgency — producing a deployable, commercially viable solution rather than a research project awaiting further development.

Young AI Leaders Community Hub

As President of this hub, Adam is building the next generation of AI practitioners committed to applying their technical skills to social and environmental challenges. The hub provides mentorship, technical training and a network connecting young innovators working at the intersection of artificial intelligence and sustainability.

Core Philosophy

Every design decision in the Smart Waste platform — from the citizen incentive system to the investment ROI model — reflects Adam's foundational principle: "Ecology should not be a punishment — but a profitable investment for everyone." Sustainability works when it is economically rational, not just morally aspirational.

Partnership Opportunities

• Cities & Municipalities — Pilot deployment agreements, data-sharing frameworks, joint EU smart city and Horizon Europe grant applications.
• Industrial Partners — Material offtake agreements for recovered resources; AI audit modules for REACH and EPR compliance via NLP processing of regulatory texts.
• Researchers & Universities — Joint publication, anonymised dataset access, co-development of next-generation sorting models and behavioural incentive designs.
• Investors — Seed and Series A discussions welcome. Full financial model, pilot performance data and technology stack documentation available under NDA.
• Technology Providers — IoT hardware, cloud infrastructure, blockchain deployment and connectivity partnerships actively sought.

Contact

📧 contact@smartwasteai.fr

📍 Toulouse, France

🌐 Follow us on Instagram & Facebook for project updates and pilot results.