Common loon on open water
Humanitarians AI  /  501(c)(3)  /  Crosslake + BostonSeptember 2026

What we will
build for the
Loon Center

Prepared for

Natasha Bartolotta

Science & Stewardship Manager, National Loon Center

The offer

We build it.
You keep it.

Humanitarians AI is a Massachusetts 501(c)(3) building AI tools for social good with graduate student volunteers. We are already active on the Whitefish Chain.

We build at no cost. We ask only for access and expertise, not a budget.

No cost

Volunteer-built. No invoice, no license fee, no subscription.

No lock-in

Tools, footage and game builds handed over to the Center. Yours to keep and run.

Already underway

Field season active now, and four game chapters playable today.

Proof, not promise

This is already
playable.

Deep Dive puts a loon family in your hands. Feed the chicks as water clarity worsens and hazards close in. One minute makes the biology consequential.

Play Deep Dive
Released
11 Sept 2026
Platform
Browser / HTML5
Session
About a minute
Built by
Shesh Narayan Gupta

No install, no sound required. Touch, mouse or keyboard. Developed from concepts in Generative AI for Game Development: Crafting Narrative Worlds with Machines.

Deep Dive title screen
A Loon's Lake  /  field notes

The game,
in play.

Each chapter is a single mechanic carrying a single fact. Gentle and challenge modes, no sound needed, and a motion toggle for visitors who need one.

Nest Site  ·  spring

Explore three shorelines and put your choice to the lake test. Open shore, high bank or sheltered cove. The brief is the real one: water access, shelter and a quiet shoreline.

Runway  ·  fall

Build lift across open water, steering around boats and docks. Headwind helps in gentle mode. A quarter mile of clear water or no takeoff.

Play any chapter now

Play Nest Site Play Deep Dive Play Chick Taxi Play Runway
Nest Site gameplay: three numbered nesting spots along a lake shoreline
Nest Site · open shore, high bank, sheltered cove
Runway gameplay: loon building lift across open water past a boat
Runway · lift, lanes, boat traffic
What we are doing on the water

Field capture:
Whitefish Chain

Drone

Aerial imagery of nesting sites, shoreline habitat and pre-migration rafts. FAA Part 107, two team members certifying now. Fall staging runs through late October, so the raft window closes with the migration.

Scuba

Underwater documentation of habitat structure and aquatic vegetation, beginning at ice-out next spring. Working toward AAUS Scientific Diver in the meantime.

Acoustic

Season-long audio recording. No certification required, so collection is already running.

Disturbance protocol

Flight logs flag operations below 50m AGL. Your guidance gets written into the tooling, not added afterward.

Common loon on the Whitefish Chain
Whitefish Chain, breeding season
Beyond the research

Footage you
can use.

We fly the Chain all season anyway. The imagery does not have to stay inside the research. It is handed to the Center at no cost, cleared for your own use.

Seasonal b-roll

Aerial coverage from ice-out through fall. Use it for exhibits, web, social, press kits and grant applications.

Construction aerials

Repeat flights from consistent positions create time-lapse and opening-day comparisons, augmenting your builder's footage.

Aerial view of a loon on summer water beside a wooded shoreline
Summer
Loon on still water against autumn color
Fall
AI tools for the Center

Four tools. All yours.

1

Metadata & annotation

Auto-tags GPS, timestamp and weather. Volunteers label species, behavior and habitat to a Darwin Core schema.

Building now
2

Acoustic call classifier

Identifies wail, tremolo, yodel and hoot calls. Male yodels are individually distinct.

Building now
3

Loon detection & counting

Detects and counts loons in drone imagery for peak-season raft counts.

After fall capture
4

Habitat change detection

Flags milfoil and pondweed; tracks shoreline and wetland change across repeat visits.

Winter – spring
The game we will make for you

A year on
the Chain

One short game follows a single loon through a year on the Whitefish Chain. Each chapter plays alone in about ninety seconds, or end to end as one story. Facts become consequences.

Adult loon swimming with a chick
Chicks ride on their parents' backs

Spring  ·  Nest Site

Loons can barely walk. Place the nest too high and the bird cannot reach it; too low and a wake swamps it.

Playable →

Summer  ·  Deep Dive

Feed chicks as water clarity drops, with lead tackle as the hazard.

Playable →

Summer  ·  Chick Taxi

Chicks ride on their parents' backs. Collect warmth, steer around wakes. The youngest visitors start here.

Playable →

Fall  ·  Runway

A quarter mile of open water to get airborne. The whole year rides on one takeoff.

Playable →

Late summer  ·  Count the Raft

Count loons in real drone footage from your lakes, then see what the model counted. The one chapter still to build, and the one that runs on Whitefish Chain footage.

The chapter only you can have

Your visitors become
your researchers.

1

Visitor counts

A real Whitefish Chain drone frame appears on the kiosk. The visitor counts the loons they can see.

On the floor
2

Model counts

The screen reveals model and true counts, and rafting behavior is explained in one screen.

Same screen
3

The Center gains data

Every human count becomes validation data. Floor traffic measurably improves the research.

Ongoing
4

Why it cannot be copied

The chapter runs on the Center's lakes and birds only. No other organization can ship it.

Yours only
Where it lives

One build. Four places.

Kiosk  ·  exhibit floor

The game chapters themselves, running in a browser on any touchscreen. No installer, app store or IT project. Each chapter is a standalone ninety-second play, so a visitor can take one and move on.

Phone  ·  on the water

Same build, same link, for visitors watching real loons from a dock and for winter crowds.

Classroom  ·  school groups

Free teacher-facing use before or after a visit extends education beyond the building.

Website  ·  before opening

The same chapters hosted on a Humanitarians AI project subdomain and embedded on your site, playable months before doors open in spring 2027.

Schedule to opening

Built before you open.

●

Sept – Oct 2026

Crosslake field season. Drone flights and acoustic recording through fall staging, all footage tagged. Capture closes when the birds leave.

Field · now
●

Nov – Dec 2026

Detection model trained on the season's raft footage. Game architecture and visual direction locked.

Build
●

Jan – Mar 2027

Volunteer chapter builds: Count the Raft joins the four chapters already playable.

Build
●

April 2027

End-to-end build delivered, staff-reviewed and ready for the floor.

Review
●

Spring 2027

Kiosk deployment and public link live for opening. Scuba documentation begins at ice-out, as the loons return.

Opening
What we need from you

Four asks.
None of them money.

Site access

Coordinate flight and dive windows, construction clearance, and which nesting areas are off limits.

Your expertise

A review pass on every fact in the game. No loon biology ships without that check.

Disturbance guidance

Approach-distance and altitude protocols built into the tooling, not worked around.

A decision, later

No commitment today. In spring, decide whether the playable build belongs on your floor.

Tool specification

Detection tool:
from footage to findings.

One tool to detect loons, invasive vegetation, shoreline change and underwater habitat structure from raw footage.

Inputs

Raw drone footage and action cameras such as GoPro for underwater and surface video, including audio tracks where available.

Core detections

  • Detect loon calls from audio
  • Count loons from aerial footage
  • Detect and locate invasive vegetation
  • Track habitat change across captures
  • Track shoreline erosion over time
  • Classify underwater footage by structure

Privacy

Blur sensitive information such as faces and license plates in footage before anything is shared or published.

Output

A report flagging footage by detected feature, with per-detection detail.

  • Loons: file, timestamp, count
  • Vegetation: file, timestamp, species
  • Erosion: file, timestamp, percentage
  • Habitat change: file, timestamps, metric
  • Underwater: file, timestamp, classification
  • Blurring: processed video files
How we get there

Build the foundation.

Assumptions

  • Users are biologists and researchers, not software engineers or ML practitioners
  • Drone footage is the footage we will have

Challenges

  • Change detection needs the same location captured at different times and spatially aligned, so real change separates from differences in drone position, angle, lighting and season
  • Training data volume is unclear; loons, vegetation species and underwater structures may each need their own labeled set
  • Ground truth needs defined owners, especially for erosion percentage and vegetation species, which require expert judgment
Plan of action

Then scale.

Phase 1
Foundation

Annotation tool converting raw footage into labeled, trainable datasets.

Phase 2
Single frame

Blur sensitive information, count loons per file, detect invasive vegetation with species and location.

Phase 3
Change

Track erosion and habitat change using aligned multi-time footage.

Phase 4
Underwater

Classify scuba footage by habitat structure.

Non-functional requirements

  • Usability: operable by non-technical users with minimal training
  • Offline capability: core detection runs without an internet connection
  • Resource efficiency: runs within modest memory limits on standard hardware
Source material

Loon-bird
image library

Annotated contact sheets from the Whitefish Chain and archive sources, labeled with species and detection confidence.

Real frames, not a demo set

This is the training set the counting model is built on, and the evidence that the tooling is already running against footage from your lakes rather than a public benchmark dataset.

Every frame is credited

Source, capture device and date travel with each file through the annotation tool, so any figure in a report can be traced back to the image it came from.

Source material  /  contact sheets
Contact sheet of annotated loon photographs with detection boxes and confidence scores
Contact sheet 01
Second contact sheet of annotated loon photographs
Contact sheet 02
Interface archive

Gavia

The field-facing front end. Upload or take a photo, and the detector says whether a loon is present and how confident it is. Built for people doing careful conservation work, not for engineers.

Honest by design

Confidence is shown as a number, every result stays reviewable, and the interface states plainly that AI results can contain errors and should be checked when accuracy matters.

Gavia: is this a loon, with photo upload
Check an image
Gavia about screen
About
Gavia previous checks list
Previous checks
Gavia result with detected loon and confidence
Detection result
Common loon on the Whitefish Chain at dusk
Four chapters done. One to go.

Play it
first.

Play Deep Dive

Nina Harris

Co-founder & Creative Director, Humanitarians AI
Crosslake, MN  |  Boston, MA

Shesh Narayan Gupta

Board of Directors, Humanitarians AI
Chicago, IL

Common loon · Whitefish Chain
A Loon's Lake by Shesh Narayan Gupta
Gavia · a product of Humanitarians AI