AI & computer vision · Research platform

Turning underwater video into a usable scientific workflow.

A browser-based platform for uploading underwater footage, running marine-life detection, and reviewing results without requiring researchers to operate the model pipeline directly.

Role: Project lead and lead developer
Status: Research
Date: 2025–Present
Organization: University of Rhode Island AI Lab

Visit public project

The labeling workspace combines AI-assisted masks with hands-on review and classification.

Product walkthrough

The working application, not just the landing page.

Key screens show how the product supports the full workflow from setup through review and action.

OceanDetect Run Insights overview with annotated underwater video and export controls — Run Insights brings annotated video, evidence review, and exports into one workflow.

OceanDetect Run Insights statistics with tracked individuals and species evidence — Statistics distinguish tracked individuals from raw detection events.

OceanDetect detections table showing species, confidence, track ID, and frame — Frame-level detections remain available for verification and quality control.

OceanDetect confidence distribution and taxon occurrence charts — Confidence and occurrence views make the model output inspectable over time.

OceanDetect sampled-frame occupancy and encounter event rate charts — Ecological summaries translate detections into reviewable time-based evidence.

OceanDetect export workspace with labeling, analytics, training, and Drive options — One export surface supports labeling, model training, scientific analysis, and shared storage.

OceanDetect Start New Run screen with model, tracker, quality, and processing options — A guided setup exposes meaningful model and processing choices without requiring a command line.

OceanDetect labeling pipeline with waiting, labeling, review, and collection stages — The labeling board tracks datasets from upload through annotation, review, and collection.

Executive summary

The operational problem

Marine researchers need to review large volumes of BRUV footage. Running computer-vision tooling manually adds setup, processing, and result-review overhead to an already demanding research workflow.

Context & constraints

What shaped the solution.

Large video files and long-running GPU inference jobs
Research workflows that need transparent, reviewable detections
Separate API, inference, export, queue, storage, and logging concerns
A product experience usable by researchers rather than model engineers

System view

A workflow built around reviewable progress.

Input

Orchestrate

Process

Review

The diagram intentionally communicates responsibility and data movement without exposing sensitive topology. Each stage has a clear user-visible state and an operational owner.

Engineering decisions

Decisions that mattered.

Designed a run-based workflow around upload, processing, review, and export
Separated GPU inference from CPU-oriented API and export responsibilities
Added persisted job state, cancellation, shared storage, and operational logs
Kept model results reviewable instead of presenting AI output as unquestioned truth

Technical depth and delivered system

Responsive browser application
Video import and run management
Detection review and scientific export workflow
Containerized service architecture

Results

What can be stated responsibly.

Researchers can begin an analysis run through a guided browser workflow.
The platform makes model processing and result review part of one product surface.
Performance and model-quality claims remain intentionally unpublished here until their evaluation records are approved.

No quantitative metric is published without an approved baseline, evaluation method, date, and caveat.

Role & collaboration

Ownership with clear attribution.

Built with student, faculty, and research collaborators at the URI AI Lab. Individual attribution will be expanded with publication approval.

Lessons & next steps

What carries forward.

A useful research platform must make uncertainty, job state, and review visible. Model performance alone does not create a dependable scientific workflow.

Next: Continue strengthening labeling, QA, export, observability, and validated model reporting.

Related work

FlowTask

A role-aware platform that brings service orders, vendors, customers, locations, documents, assets, and operational search into one coherent workflow.

Read related case study