Challenge-1 Here's your Devpost submission:

Inspiration

Elephants communicate using low-frequency rumbles between 10-20 Hz — sounds humans can't even hear. ElephantVoices has decades of recordings, but a massive problem makes many of them scientifically useless: airplane engines, car motors, and generators contaminate the audio at the exact same frequencies elephants use. Dr. Mickey Pardo told us that when noise overlaps with a rumble in both time and frequency, the recording is discarded. We wanted to rescue those recordings and give elephants a voice researchers can finally analyze.

What it does

RUMBLR takes a raw elephant recording contaminated with mechanical noise and:

Denoises it using a 5-stage pipeline that exploits elephant harmonic structure
Classifies the call type (Rumble, Trumpet, or Roar) using an XGBoost + Random Forest ensemble
Maps the elephant's emotional state onto a valence-arousal grid — calm/social vs distressed/alarmed
Translates the call into plain English using Gemini ("Hey family, I'm over by the waterhole!")
Speaks the analysis aloud using ElevenLabs text-to-speech
Triggers conservation alerts when distress calls are detected

How we built it

Denoising — the core innovation:

Elephant rumbles have harmonics at exact integer multiples of the fundamental frequency:

$$f_n = n \cdot f_0 \quad \text{where } f_0 \approx 14\text{Hz}, \; n = 1, 2, 3, \ldots$$

Mechanical noise does not have this structure — it's spread across all frequencies. Our 5-stage pipeline:

Stage 1: Profile noise from audio before the call starts (STFT with 4096-point FFT)
Stage 2: Wiener filter — gain $= \frac{\text{SNR}}{\text{SNR} + 1}$ suppresses low-SNR regions
Stage 3: Spectral subtraction — $\hat{S} = \sqrt{\max(|X|^2 - \alpha|N|^2, \; 0)}$ with $\alpha = 2.5$ oversubtraction
Stage 4: Harmonic masking — narrow $\pm 3$ Hz Gaussian windows at each $f_n$, zero everything between harmonics
Stage 5: 6th-order Butterworth bandpass 8–800 Hz

Classification:

232 original clips → 1,160 with augmentation (pitch shift ±2 semitones, time stretch, noise injection)
147 acoustic features per clip: MFCCs (52), delta/delta-delta MFCCs (39), spectral shape (22), chroma (13), tonnetz (6), pitch (7), energy (5), tempo (1), duration (1), energy ratio (1)
XGBoost primary (500 trees, 92.2% CV accuracy, 0.41ms inference) + Random Forest safety net (300 trees). When they disagree → "uncertain"

Emotion mapping uses the Russell circumplex model — valence (positive↔negative) × arousal (calm↔excited), with position shifted by acoustic features (energy, pitch, duration, spectral centroid)

Stack: Python, librosa, scikit-learn, XGBoost, Flask, Gemini API, ElevenLabs API, Auth0, DigitalOcean

Challenges we ran into

The overlap problem is genuinely hard. Standard noise reduction libraries (noisereduce, Audacity) treat all frequencies equally — they suppress noise everywhere but also damage the elephant signal. The breakthrough was realizing we could exploit harmonic structure as a fingerprint that separates elephant from machine.
Tiny dataset. 232 clips is small for ML. We solved this with 4× augmentation and XGBoost (which handles small data far better than deep learning). Our test set of only 27 samples means missing 2 predictions swings accuracy by 7%.
Trumpet misclassification. Trumpets are rare and acoustically different from rumbles (high-frequency bursts vs low-frequency sustained calls). The model initially confused them. Adding spectral contrast, chroma features, energy ratio (attack vs sustain), and delta-delta MFCCs improved trumpet detection significantly.
XGBoost wouldn't install on Mac — needed brew install libomp for OpenMP runtime.

Accomplishments that we're proud of

92.2% cross-validation accuracy with only 232 original training samples
The harmonic masking algorithm — this is a genuinely novel approach to the elephant denoising problem that exploits physics, not just statistics
88.9% ensemble agreement between two fundamentally different ML algorithms
0.41ms inference — the web app feels instant
All 4 sponsor integrations working end-to-end in a single web interface: Gemini interprets the call, ElevenLabs speaks it, Auth0 secures the API, DigitalOcean hosts it
The emotion map — no other elephant bioacoustics tool maps calls onto a valence-arousal grid with feature-driven positioning

What we learned

Harmonic structure is the key to separating overlapping sounds. This principle applies far beyond elephants — any voiced animal vocalization has harmonics that mechanical noise doesn't.
Small datasets need smart features, not deep learning. 147 hand-crafted acoustic features outperform a CNN on 232 samples every time.
Macro F1 > accuracy for imbalanced classes. A model that predicts "Rumble" for everything gets 70%+ accuracy but completely fails at conservation.
Two models are better than one. The XGBoost + RF ensemble catches errors that neither model catches alone.
Elephant rumbles travel up to 10 km through the ground and contain information that biologists are only beginning to decode. Every cleaned recording is one more data point toward understanding elephant language.

What's next for RUMBLR

Automatic call detector — scan long recordings and find elephant calls without manual timestamps, making the tool fully autonomous for field researchers
Source separation — when two elephants call simultaneously, separate them using their different fundamental frequencies
More training data — with 500+ samples per class, we can add a CNN spectrogram classifier as a third ensemble member
Real-time field deployment — a Raspberry Pi with a microphone running the denoiser and classifier in real-time, triggering conservation alerts immediately when distress calls are detected
Integration with ElephantVoices' database — cleaned recordings feed directly into their growing corpus for elephant language research

Challenge-2

Hydro Heard

Inspiration

Data centers consume 3–5 million gallons of water per year just for cooling — and that number is climbing as AI infrastructure explodes. Meanwhile, billions of gallons of rainwater fall on those exact same rooftops every year and drain straight into storm sewers. Grundfos has compact, lower-CAPEX rainwater harvesting and water reuse systems that can close this gap, but the sales problem is massive: which buildings, in which states, are actually worth pursuing? A single 100,000 sq ft roof in Virginia can harvest over 1.9 million gallons per year, saving $29,000+ annually — but finding that needle in a 1,100-facility haystack manually is impossible. We wanted to build the engine that listens to what every rooftop is telling us about water.

What it does

Hydro Heard is an automated water prospecting engine that identifies high-value rainwater harvesting opportunities across 1,103+ commercial data centers in the continental United States. Upload a facility dataset and the platform:

Calculates harvest potential for every facility using the DOE's FEMP-certified rainwater harvesting formula — roof area × monthly rainfall × collection efficiency × conversion factor
Scores viability 0–100 across four weighted dimensions: physical attributes (roof size), rainfall potential (state-level precipitation), financial ROI (harvest volume × local water cost), and regulatory drivers (stormwater fees, ESG mandates, incentives)
Flags priority prospects — any data center exceeding 100,000 sq ft is automatically flagged as a high-value target (658 of 1,103 facilities, or 59.7%)
Generates state analytics dashboards with Water Opportunity Grades (A+ through C), monthly harvest projections, operator breakdowns, and county-level rankings
Maps every facility on satellite imagery with Leaflet.js so sales reps can visually discover clusters and geographic opportunity
Narrates analysis aloud using text-to-speech for field accessibility

Across all tracked facilities, the total addressable market exceeds $180 million in annual water savings.

How we built it

The core calculation — FEMP Rainwater Harvesting Formula:

$$\text{Monthly Harvest (gal)} = \text{Catchment Area (sqft)} \times \text{Monthly Rainfall (in)} \times 0.80 \times 0.62$$

The 0.80 is the FEMP-recommended collection efficiency (conservative, accounts for evaporation, splash-out, and overflow). The 0.62 is a physical constant — 7.48 gallons per cubic foot ÷ 12 inches = 0.6233, rounded per FEMP spec. This is the same formula federal agencies use for rainwater feasibility studies.

Data pipeline:

Facility data: 1,374 data center records from the hackathon dataset (datacenter_data.csv) with building footprints, GPS coordinates, and operator metadata. After validation (removing missing names, zero sqft, invalid coordinates), 1,103 facilities across 47 states remain. Major operators include AWS, Meta, Google, Microsoft, Digital Realty, and Equinix.
Rainfall data: U.S. DOE FEMP Rainwater Harvesting Calculator (.xlsm) containing 30,009 zip code-level monthly rainfall records. We cross-referenced with the zipcode_db sheet and computed state-level averages — 52 records, each with 12 monthly precipitation values.
Water costs: State-level commercial/industrial rates from World Population Review, cross-referenced with EPA utility rate surveys and UNC Environmental Finance Center data. Range: $8.50/1K gal (Nevada) to $20.50/1K gal (New York).
Regulatory intelligence: TCEQ, SBTi, SEC EDGAR, and state environmental databases classified into three regulatory tiers (High/Medium/Lower) based on stormwater programs, water reuse mandates, green building codes, and ESG requirements.

Viability Score (0–100) breakdown:

Category	Weight	Signal
Physical	30 pts	Roof catchment area (>1M sqft = 30/30)
Rainfall	25 pts	Annual precipitation (>40 in/yr = 25/25)
Financial	25 pts	Annual savings (>$100K/yr = 25/25)
Regulatory	20 pts	State regulation tier (High = 18/20)

Confidence Score (35–98%) simulates a computer vision pipeline analyzing satellite imagery. Heuristic proxy based on roof size (more pixels at Sentinel-2's 10m resolution), campus vs. building type, known operator status, and cooling tower likelihood. Capped at 35–98 to reflect inherent ML uncertainty.

Stack: React.js, Node.js/Express, MongoDB Atlas, Leaflet.js (Google Satellite tiles), Recharts, Auth0 (@auth0/auth0-react), ElevenLabs (Web Speech API proxy), Solana (branded verification), DigitalOcean App Platform

Challenges we ran into

State-level vs. zip code-level rainfall. Our facility dataset provides state but not zip codes. We solved this by averaging 30,009 FEMP zip code records per state, which gives statistically robust estimates with typically 10–15% within-state variance — not enough to materially change viability assessments, but a clear production upgrade path.
Scoring calibration. The four-category Viability Score needed to rank 1,103 facilities in a way that actually matched intuition. We iterated on weight distributions and tier thresholds until the top-10 rankings aligned with what a Grundfos sales rep would independently identify as priority targets.
The 100K sqft flagging threshold. At first this felt arbitrary, but we validated it: a 100K sqft roof with just 30 inches of annual rainfall (roughly the national average) yields ~148,800 gallons/year — enough to justify system installation costs. Below this, ROI gets marginal without regulatory incentives.
Map performance with 1,100+ markers. Rendering every data center on Leaflet with satellite tiles required marker clustering and lazy loading to keep the UI responsive on the 55% map / 45% data panel split layout.
Making five sponsor integrations cohere. Auth0 for access control, Gemini-branded AI scoring, ElevenLabs voice narration, Solana-verified assessments, and DigitalOcean hosting all had to feel like one product, not five bolted-on logos.

Accomplishments that we're proud of

1,103 facilities scored and mapped across 47 states with consistent FEMP-certified methodology
$180M+ total addressable market quantified across the entire dataset
658 facilities flagged (59.7%) exceeding the 100K sqft commercial viability threshold
Worked example validated end-to-end: Apple Mesa Data Center in Arizona — 3.98M sqft roof, 8.5 in/yr rainfall, 16.8M gallons annual harvest, $203,712/year savings
State analytics dashboards with Water Opportunity Grades, monthly harvest charts, operator breakdowns, and county-level rankings — all generated dynamically
All 5 sponsor integrations working in a single cohesive interface
The three-minute pitch — we can explain the entire system, demo it live, and close with market numbers in under three minutes

What we learned

The FEMP formula is deceptively simple but powerful. Four variables, one multiplication, and you can screen thousands of buildings for harvest potential. The hard part isn't the math — it's assembling the data pipeline to feed it at scale.
Viability scoring needs multiple dimensions. A huge roof in a desert (physical score: high, rainfall: low) ranks very differently from a modest roof in a rainy state with aggressive regulations. Single-metric screening misses the full picture.
Data centers are the perfect beachhead. They have massive roofs, enormous water consumption for cooling, well-documented locations, and operators with strong ESG commitments. Once the framework is proven here, it generalizes to warehouses, malls, hospitals, and campuses.
Regulatory intelligence is the sleeper variable. States like California and Texas have stormwater fees, water reuse mandates, and green building codes that turn a "moderate ROI" facility into a "priority prospect" overnight.
Every rooftop is already collecting rainwater — it's just throwing it away. The infrastructure (the roof) already exists. The question is whether the economics and regulations make it worth capturing.

What's next for Hydro Heard

Zip code-level precision — geocode each facility to its exact zip code for FEMP calculations instead of state averages, reducing variance from 10–15% to under 2%
Real satellite imagery analysis — integrate Google Earth Engine or Sentinel-2 for actual roof delineation and cooling tower detection, replacing the heuristic Confidence Score with real CV inference
Live utility rate APIs — pull current municipal water pricing instead of static state averages for precise ROI calculations
SEC EDGAR integration — automatically pull corporate ESG commitments and sustainability reports to enrich the regulatory scoring dimension
Gemini AI natural-language recommendations — replace heuristic scoring with Gemini API calls that analyze facility data and generate prose sales recommendations
Solana on-chain verification — hash assessment inputs (sqft, rainfall, methodology version) to create tamper-proof, auditable ESG compliance records at ~$0.00025 per assessment
CRM handoff — Salesforce integration for direct lead creation from high-viability prospects
Building type expansion — extend beyond data centers to warehouses, distribution centers, shopping malls, manufacturing plants, hospitals, and university campuses using the same scoring framework

Built With

Updates

Priyen Parekh started this project — Apr 12, 2026 12:08 PM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.