Rockjumper Birding AI

Find. ID. Learn.

Audio recognition, image recognition, and predictive intelligence for bird apps.

On-device inference. Regional depth. Built for partners.

FINDPredictive range maps and site lists IDPhoto and audio recognition at plumage level LEARNSimilarity-driven quizzes and gap analysis

Powering bird apps on four continents

Collins Bird Guide

Roberts Birds of Southern Africa

Sasol eBirds Southern Africa

BirdPro

In development

Sibley Birds North America

Morcombe's Birds of Australia

Stewart Australian Bird Calls

01 — FIND

Range Maps, Site Lists & Predictive Intelligence

Know where every species is, right now — driven by detection probabilities conditioned on occurrence, habitat, season, weather, and time of day. Built on an AUCⓘ of over 97% in Southern Africa and 96% in Europe (available globally).

Non-Seasonal Species

For a river specialist like the African Fish Eagle, habitat-aware maps reveal the river networks — not smeared across coarse grid cells.

Typical range map — noisy

Rockjumper — habitat-aware

Zambezi / Okavango / Orange
River networks highlighted at resolution

"Detection probabilities resolve to habitat, not grid cells. For a river specialist like the African Fish Eagle, that difference is the difference between a useful map and a guess."

Animated Seasonal Maps

The same detection model produces both a snapshot for today, and the movement across the year.

Common Quail follows rainfall. Winter rains in the Cape draw them south; summer rains on the highveld pull them north. The model captures the movement because it's conditioned on weather, not just occurrence history.

Species Page — Anatomy

Every stat is derived from the same detection model — no hand-curated tables, no stale data.

Summary stats

Peak Findability

How hard/easy the species is to detect at its best

Tick Difficulty

Composite score for twitchers adding to a life list

Activity charts

Daily activity (hour-by-hour) and seasonal variation (month-by-month) — both derived from the model.

Site Species Lists

The model doesn't need prior visits. It reads the habitat, the season, and the weather, and predicts what the birder is likely to see.

Willow Lakes

Popular site near London

Habitat context

Rockjumper — habitat & season-aware

Expected species count gives a realistic target for a single trip, not the all-time total.

Out-of-season species (Redwing) are correctly suppressed. In-season species (Blackcap) get a Peak tag.

Haven Point

Coastal site east of London — sparsely visited

Habitat context

Rockjumper — coastal-aware from day one

Shore and seabirds — gulls, oystercatchers, shelduck — surfaced first, because the model knows this is a coastal site.

A year-round list is available from day one — no cold-start problem.

02 — ID

Photo & Audio Recognition

Identify birds to plumage level from photos, and to species from calls — even in challenging conditions.

Plumage-Level Photo ID

Species-level is a floor, not a ceiling. The model goes to plumage — telling the birder what they actually saw.

Step 1 — Input

User's photo — a small brown bird on a lawn.

Step 2 — Auto-crop

The model locates and crops the bird before classification — no manual framing required.

Step 3 — Result

European Robin — juvenile plumage — 98% confidence.

Most competitors return: European Robin — adult.

Competitor result showing adult European Robin

The difference between adult and juvenileEuropean Robin is the difference between a bird a birder has seen a thousand times and one they've probably never knowingly identified. Plumage-level matters.

Call ID

Call-level recognition — not just species-level — so every detection gives a specific call type the user can replay and learn from.

1.Results show a comparison spectrogram for immediate visual confirmation.

2.Confidence increases as it listens.

3.Easy playback of your recording and the library call.

Coverage & Validation

Deep regional coverage where global models fall short.

Southern Africa

Rockjumper67% of non vagrants

Merlin17% of 'likely birds'

Deep regional coverage where global models fall short.

Australia

RockjumperSignificantly broader

Merlin40% of 'likely birds'

Merlin's Australian pack is narrow; Rockjumper has significantly broader coverage.

Europe & North America

RockjumperParity + call-type depth

MerlinSpecies-level only

Parity or better at species level, plus call-type granularity the global models don't attempt.

03 — LEARN

Similarity, Quizzes & Targeted Learning

Help users learn by surfacing the species they're most likely to confuse — and by pointing them at the species they're statistically missing.

Similar Sounds

For each species, the similarity graph surfaces the acoustically closest matches. Only matches above 40% shown.

Primary

Spotted Crake

alarm call

0:00 / 0:00

Similar species

Common Crane

song

67%

0:00 / 0:00

Boreal Owl

flight call

58%

0:00 / 0:00

Eurasian Skylark

alarm call

45%

0:00 / 0:00

Corn Crake

call

40%

0:00 / 0:00

Primary

Coal Tit

song

0:00 / 0:00

Similar species

Eurasian Hoopoe

song

51%

0:00 / 0:00

African Blue Tit

song

44%

0:00 / 0:00

Similar Plumages

The same approach applied to images. Only matches above 40% shown.

Primary

Surf Scoter

adult female

most similar →

Velvet Scoter

adult female

63%

Harlequin Duck

female

58%

Common Scoter

juvenile

56%

White-headed Duck

female

54%

Common Scoter

adult male

50%

Ruddy Duck

adult breeding male

47%

Greater Scaup

non-breeding female

47%

Ruddy Duck

adult female

47%

Velvet Scoter

adult male

45%

Bufflehead

female

43%

Primary

European Robin

adult

most similar →

Red-flanked Bluetail

non-breeding

71%

Red-breasted Flycatcher

adult male

65%

Red-flanked Bluetail

adult male

56%

Moussier's Redstart

female

56%

Thrush Nightingale

non-breeding

47%

Common Redstart

female

44%

Veery

adult

43%

White-throated Robin

adult male

41%

Expert Call Quiz

Similarity isn't just a data structure — it's the engine for adaptive quizzes. The closer two species are in the similarity graph, the harder the pairing. Novice mode pairs species from different families; expert mode pairs species the model itself finds nearly identical.

1.Two acoustically similar species, drawn from the similarity graph

2.User drags the species chip onto the call that matches

3.Difficulty scales: closer pairs = harder quiz

User Gap Analysis & Targeted Learning

Every birder has blind spots — species they pass by, or mis-assign, or never learn to separate from a close cousin. Gap analysis finds those blind spots per-user, per-site, using the detection model as ground truth. Then the learning content targets them specifically.

Diagnose

User lists

Per-site observations

→

SDM predictions

Expected detections

→

Gap report

Under-detected taxa

e.g. “Larks — 3 spp. missing”

Learn

Habitat guidance

Where to look

→

ID comparison

Similar plumages & calls

→

AI quiz

Generated pairs

Next outing triggered

Apply

Where & when

Site, habitat, time of day

→

Target species

Most likely gaps here

→

Field ID sheet

Side-by-side marks

↺ Lists updated — cycle repeats

04 — CONTENT

Automated Classification for Guide & Library Builders

Beyond user-facing features — tools for building and enriching bird-guide content at scale.

Automated Call-Type Classification

Similarity models automatically cluster recordings by call type. Labels are human-assigned after grouping; the hard work — grouping 150 recordings into 3 coherent call types — is done by the model.

Point it at an unlabelled recording archive and get back structured call-type groups.

Eurasian Jay

Group 126%

Eurasian Jay — Group 1

0:00 / 0:00

Group 223%

Eurasian Jay — Group 2

0:00 / 0:00

Group 352%

Eurasian Jay — Group 3

0:00 / 0:00

Thrush Nightingale

Group 151%

Thrush Nightingale — Group 1

0:00 / 0:00

Group 219%

Thrush Nightingale — Group 2

0:00 / 0:00

Group 330%

Thrush Nightingale — Group 3

0:00 / 0:00

Automated Plumage Classification

The same similarity logic applied to images — automatically grouping photos by plumage, age class, or morph. A guide publisher can point it at 500 unlabelled photos and get back coherent plumage groups.

Useful for field guides that want to show the bird as birders actually see it.

European Robin

Juvenile8%

Adult92%

The model finds the split without being told what to look for.

Hen Harrier

Perched8%

Perched — female/juvenile24%

Flying22%

Flying — female/juvenile46%

Not just plumage — posture and behaviour fall out of the grouping too.

05 — HOW IT WORKS

The Engine

The models that power the features above — not black boxes.

>97% AUC

Southern Africa

Detection Modelⓘ

Site-level species predictions conditioned on habitat, season, weather, and time of day. >96% AUC in Europe. Site-level predictions coming globally.

Call-type level

Not just species

Audio Recognition

Outputs snap to a specific call the user can replay. Accuracy is boosted further by feeding in local species likelihoods from the detection model.

Plumage-level

Age, sex, morph

Image Recognition

Tells the user what they actually saw, including age class, sex, and morph where relevant — not just the species name.

On-device

No signal needed

Deployment

Audio and image models run on-device — available out of signal range, no server round-trip, no ongoing infra cost. Detection-model outputs are precomputed and cached per region, updated periodically.

6 model types

Working together

Model Ensemble

Seven specialised image and audio models work in concert to produce image and audio identification, similarity and grouping