REWILD

Inspiration

Every spring, millions of homeowners Google the same question: "How do I help pollinators in my yard?" They find static plant lists, pretty mockups from AI landscape tools, and research simulators built for PhDs — but nothing that answers the real question: "If I plant a native meadow instead of a rain garden, what actually happens to the butterflies by Year 3?"

We were inspired by three colliding trends:

The native gardening movement is exploding. 83 million US households have yards, and a growing wave of homeowners are ripping out lawns and planting native species. But they're flying blind — choosing interventions based on blog posts and Instagram aesthetics, not ecological outcomes.
Uncertainty is the honest answer — and no one shows it. In ecology, projecting what happens over 5 years is inherently uncertain. Every existing tool gives you a single confident-looking answer. We wanted to build something that treats uncertainty as a feature, not a bug — showing optimistic, likely, and conservative bands so users actually understand the range of possible futures.
Ecological simulation is locked behind institutional walls. Tools like LANDIS-II and InVEST are powerful, but they require GIS expertise and institutional licenses. We asked: what if you could get a meaningful ecological scenario comparison from just a ZIP code and a 2-minute wizard?

REWILD was born from these frustrations — a tool that finally bridges the gap between "I want to rewild my yard" and "here's what the science says might happen."

What it does

REWILD is a consumer ecological scenario engine that lets anyone simulate and compare habitat restoration outcomes over 5 years — from a backyard to a school garden.

Here's the user journey:

Enter your ZIP code → REWILD auto-detects your USDA hardiness zone, EPA Level III ecoregion, and frost dates. No guesswork.
Describe your site → Current state (lawn, weedy lot, bare soil), sun exposure, soil type, and your goals (pollinators, birds, water management, carbon, beauty).
Choose interventions → Compare up to 3 at once: native meadow, rain garden, shrub border, pollinator nesting stations, stop mowing, leave leaves, habitat structures, or native grass plantings. Each one is scored for your site's specific conditions.
Watch the future unfold → Interactive trajectory charts track 4 ecological metrics — pollinator diversity, bird activity, food-web complexity, and ecosystem services — over 5 years. Every single data point shows optimistic/likely/conservative bands with confidence percentages that decay over time. Toggle between overlay and side-by-side views.
See the food web grow → An animated trophic network visualization shows your ecosystem expanding year by year from bare soil to a complex web of plants → pollinators → birds. Watch Monarchs arrive when milkweed establishes in Year 2, hummingbirds appear when cardinal flowers bloom in Year 3.
Get your action plan → A frost-date-aware planting calendar, species shopping list (with quantities calculated from your area), 8-week site prep timeline, seasonal maintenance checklists, and a "Top 10 Mistakes to Avoid" guide. Print the whole thing and take it to the garden center.

What makes it different from everything else:

It's a scenario comparison engine, not a plant finder. The core question is "what happens differently?" not "what should I plant?"
Non-linear Ecological Synergy: Most calculators just add scores together. REWILD computes pairwise mathematical synergies when you combine interventions. If you plant a Native Meadow and install Pollinator Nesting Stations, the model calculates a 1.62x non-linear boost to pollinator trajectories because the interventions support different life stages.
It's uncertainty-first — every projection comes with confidence bands and actionable suggestions to narrow them ("Add a soil test to narrow ecosystem service predictions by ~15%").
It builds a trophic food web from real species data, not decoration. You see which specific pollinators arrive based on which plants have established.

How we built it

Architecture: React 19 + Vite frontend → FastAPI + Uvicorn backend, with a multi-layered simulation engine and AI narrative layer.

The Simulation Engine (the hard part)

We built a 5-layer simulation pipeline entirely from scratch:

Layer 1 — Deterministic Succession Model: We hand-calibrated base trajectory curves for all 8 interventions across 4 ecological metrics, then modulated each trajectory by five independent factors:

Ecoregion speed multipliers (Florida ecosystems recover 1.3× faster than Northern Forests)
USDA zone growing-season factors
Site condition modifiers (a partial garden starts at 1.25× vs bare soil at 0.85×)
Sun/shade penalties (shade plots reach only 65% of full-sun potential)
Area bonus (log-scaled — larger plots are marginally more effective)

Layer 2 — Uncertainty Propagation: Every metric value gets wrapped in a 4-component uncertainty calculation:

Ecoregion study confidence (well-studied Eastern Temperate Forests = 0.80, sparse North American Deserts = 0.60)
Intervention documentation confidence (native meadow = 0.80, leave leaves = 0.55)
Time decay (6% confidence loss per year — Year 5 predictions are inherently wider)
Soil knowledge bonus (unknown soil widens bands by 15%)

The system also generates actionable uncertainty reducers — specific suggestions like "Get a soil test to narrow plant survival predictions by ~15%."

Layer 3 — Food Web Builder: We built a trophic network generator that uses real ecoregion-specific native plant and pollinator databases. For each simulation year:

Plants establish based on intervention type and time (active plantings colonize faster than stop-mowing)
Pollinators arrive only when their specific food plants exist (no fake connections)
Bird guilds appear at food-web complexity thresholds (ground nesters need 0.4+ complexity)
Network metrics (connectance, trophic levels) are computed per year

Layer 4 — AI Narrative Engine: A two-tier system where OpenAI GPT-4o-mini generates ecological narratives explaining the "why" behind the numbers. If the API is unavailable, a template engine provides instant offline-capable fallbacks. The AI generates 2-3 paragraph stories, top 5 species recommendations, and a key ecological insight.

Layer 5 — Succession-Aware Bloom Phenology: We built a temporal bloom calendar that evaluates how evenly a garden blooms across 12 months using a normalized Shannon-entropy metric (Bloom Continuity Score). Because it is succession-aware, it models when plants establish—fast-growing forbs bloom in Year 1, while slow-growing shrubs and trees don't provide bloom coverage until Year 3+. It also analyzes temporal gaps and automatically recommends specific native species to fill "low-bloom" months.

Data Layer

We curated 5 ecological datasets from public sources:

USDA Hardiness Zones — 3-digit ZIP prefix → zone mapping for all US ZIP codes
EPA Level III Ecoregions — ZIP-to-ecoregion mapping with regional characteristics
Native Plant Database — Ecoregion-specific species with bloom seasons, ecological value, and sun/soil requirements
Pollinator Registry — Bees, butterflies, moths with conservation status, flight seasons, and plant visitation relationships
Intervention Library — 8 intervention types with suitability scores, cost ranges, and difficulty ratings

Frontend

Built with React 19 and Recharts for trajectory visualization, and a custom Canvas 2D renderer for the animated food web graph. The UI follows a 4-screen wizard-to-dashboard flow designed for progressive disclosure — users never see complexity until they need it.

Challenges we ran into

Honest uncertainty is visually messy. Showing three bands (optimistic/likely/conservative) on a chart with multiple interventions overlaid gets noisy fast. We solved this with careful opacity layering, toggle between overlay and side-by-side modes, and confidence percentage badges that give users a single number to anchor on.
Ecological data is fragmented and inconsistent. USDA zone data uses different ZIP granularity than EPA ecoregion data. Native plant databases disagree on species ranges. Pollinator-plant relationships are incompletely documented. We spent significant time normalizing and cross-referencing these datasets into a consistent format.
Calibrating succession curves without longitudinal data. No public dataset says "here's exactly how pollinator diversity changes year-by-year when you convert lawn to meadow in Zone 7a." We synthesized our base curves from restoration ecology literature, extension service publications, and Pollinator Partnership guides, then validated against documented case studies.
Soil-Aware Cross-Effects: We didn't just use soil to filter plant lists; we integrated soil type directly into the succession math. The challenge was that soil interacts with metrics and interventions differently. The engine knows that a Rain Garden built on compacted clay will perform 22% better for ecosystem services than one built on sandy soil (which drains too fast), actively changing the outcome bands dynamically.
Food web realism vs. performance. Generating a trophic network with real species interactions that grows believably over 5 years — while keeping the Canvas rendering smooth — required careful gating logic. We implemented threshold-based colonization (pollinators only arrive when their food plants exist, birds only appear at sufficient food-web complexity) rather than random population.
Making uncertainty actionable, not just decorative. Most tools show error bars and call it a day. We wanted uncertainty to teach — so we built the reducer system that tells users specifically what they can do to narrow their predictions. This was conceptually simple but required tracking which uncertainty sources are addressable.

Accomplishments that we're proud of

The food web animation is a genuine "wow" moment. Watching a trophic network grow from 5 nodes (Year 0: bare soil + a few weeds) to 38 nodes (Year 5: plants → pollinators → birds fully interconnected) — with real species names and conservation status highlighting — is something no other consumer ecology tool does. When you see "Monarch Butterfly" appear in Year 2 because milkweed has finally established, it clicks in a way that numbers never do.
The uncertainty reducers change the conversation. Instead of "this prediction might be wrong," we say "get a soil test and this prediction gets 15% tighter." That turns uncertainty from a limitation into a learning pathway. Users leave knowing what to observe, not just what to plant.
It actually works for diverse locations. Enter ZIP 75254 (Dallas, TX — Great Plains) and you get different species, different trajectories, and different confidence levels than ZIP 02138 (Cambridge, MA — Eastern Temperate Forests). The ecoregion modulation isn't cosmetic; it fundamentally changes the simulation.
Offline-capable AI layer. The fallback template engine means the entire app works without any API keys. The AI adds narrative richness, but the core simulation and all actionable outputs work completely offline. You can demo this on an airplane.
Stateless & Database-free Architecture: By cleanly curating and embedding ecological datasets (USDA zones, EPA ecoregions, species registries) directly into the API layer, the entire backend is stateless. This means infinite scalability, zero database management, and lightning-fast simulation execution.
Full pipeline — ZIP code to printable planting calendar. Most hackathon projects stop at the visualization. We built the entire pipeline through to a print-ready action plan with frost-date-aware planting schedules, species quantities calculated from the user's area, and seasonal maintenance checklists. You can literally take the output to a garden center.

What we learned

Uncertainty communication is a design problem, not a math problem. The math for confidence bands is straightforward. Making them comprehensible to non-scientists required extensive iteration on visual encoding — opacity, color, layout modes, and the decision to always show a single "confidence %" number alongside the bands.
Ecological modeling is humbling. Even simplified succession models reveal how many variables interact. A rain garden in Zone 8a responds totally differently than one in Zone 4b, and capturing that with modular multipliers was an exercise in principled approximation.
Progressive disclosure is essential for complex tools. The 4-screen wizard flow (ZIP → site → interventions → dashboard) was critical. Early prototypes that showed everything at once overwhelmed users. Each screen answers one question before introducing the next layer of complexity.
Curated data beats dynamic data for hackathons. We initially considered using live APIs for plant and pollinator data. Embedding curated datasets directly in the backend made the app faster, more reliable, and let us control data quality — a trade-off that saved days of integration work.
The gap between "academic" and "consumer" is huge. Professional ecological tools model dozens of interacting variables at fine spatial resolution. Consumer tools need to be opinionated — choosing which 4 metrics matter most, which 8 interventions to offer, and how to translate index values into language like "moderate pollinator diversity."

Viability & Target Audience

While REWILD is free for individual homeowners, the underlying engine has clear paths to sustainability:

B2B Licensing for Nurseries: Garden centers can embed REWILD on their sites to drive hyper-local, ecologically appropriate plant sales.
Municipal & School District Tiers: Pro dashboards for cities tracking their biodiversity goals across public lands and schools.
NGO Partnerships: Whitelabeling the tool for conservation groups to power their specific regional rewilding campaigns.

What's next for REWILD

Observation calibration loop: Users report real observations ("I saw a monarch today," "3 bee species on my coneflowers") that update trajectory confidence in real time — creating a citizen-science data flywheel that makes the model better for everyone.
Photo-based site assessment: Upload a photo of your yard and computer vision identifies current vegetation, estimates sun exposure from shadow patterns, and suggests soil type from visual cues — reducing the wizard to a single photo.
Community layer: See what your neighbors are planting, view aggregate rewilding impact at neighborhood/ZIP/county scale, and coordinate restoration corridors that connect isolated habitat patches.
Institutional tier: Partner with municipal governments and school districts for campus-scale rewilding planning with higher-resolution spatial data and multi-year monitoring dashboards.
Model calibration from real data: As observations accumulate, use Bayesian updating to calibrate succession curves per ecoregion — turning REWILD from a scenario comparison tool into an increasingly accurate predictive model grounded in real-world outcomes.
Expanded geography: Add support for Canadian provinces, UK/EU ecoregions, and eventually tropical zones — each requiring new species databases and succession curve calibrations.