Bishop Case Study

Back to Projects

AI-powered drone search & rescue platform designed for speed, safety, and hope.

6 months to MVP

Lead Software Engineer

Patrick Sherlund

Lead Product Designer

Shelby Reilly :)

See case study below

The Team

Patrick Sherlund

Lead Software Engineer

Shelby Reilly

Lead Product Designer

My role

Engineered Bishop's search and rescue video intelligence pipeline to automate frame-by-frame review into prioritized human detections, cutting time-to-first-lead from hours to seconds.

The Stack

FRONTEND

React, vite (Typescript, Javascript)

BACKEND

Drogon (C++)

ONNX Runtime (C++, Python)

DATA

PostgreSQL

WORKFLOW

GitHub (Agile)

INFRA

AWS, Docker

The Challenge

Locating distressed humans quickly & efficiently during search & rescue missions.

The Vision

AI powered drone search & rescue application designed for speed, safety, and hope.

Our users saw...

30% ↓

Time-to-first-find reduction

3,400×

Faster video review

01 | Discover

Understand the context, painpoints, and user insights

Why this started

We started this after the 2024 Southeast Coastal hurricanes, when SAR teams struggled to find survivors quickly across flooded zones.

Discovery interviews & workshops

Persona Mapping

What We Learned

😔

The Reality Today

In SAR, the first ~72 hours are the most critical, but manual video review is fatiguing and error prone, and even reviewed footage can hide missed detections. Bishop runs a rapid second pass to surface missed human leads faster.

What's Holding Teams Back

In SAR operations, responders have ~72 hours to find subjects, but manual video review creates cognitive fatigue that increases missed detections and delays time-to-discovery.

What Success Looks Like

SPEED

Hours to Minutes

quicker detection

ACCURACY

90%+

detection rate

TRAINING

<15 min

to get started

HARDWARE

Real-time

on standard CPU

Our proposed process

Create Project

Upload Video

Review Results

Share with Team

What we proved

Validate aerial person-detection accuracy

↓

We saw 94% Recall on footage

Benchmark CPU-only video processing throughput

↓

We processed 3,700 frames per minute

Prove smooth playback of annotated timelines

↓

Roughly 60 FPS during video playback

Concepts, ideations, iterations

Design Decisions

01 | Performance

There's tons of data,
it must be fast

We need elastic scaling
with increasing workload

Deterministic
performance with
reliable, repeatable
outcomes

02 | Seamless Integration

Interoperates with
existing infrastructure

Aligns technology
with existing user
workflows

Mirrors the existing
ecosystem and
interfaces

03 | User Experience

Understands quickly
and acts immediately

Low friction use with
minimal training

Adopts quickly
across teams

Iteration 1

Week 2 - Gather, Annotate & Train

Gather datasets from published research

Annotate segmentations within the images

Train & fine tune the model

Plotting labels to /home/patrick/ml/bishop/runs/y12_phase4_mixed20k_ms102/labels.jpg...
optimizer: AdamW(lr=0.001, momentum=0.937) with parameter groups 113 weight(decay=0.0), 120 weight(decay=0.0006000000000000001), 119 bias(decay=0.0)
Resuming training /home/patrick/ml/bishop/runs/y12_phase4_mixed20k_ms10/weights/best.pt from epoch 10 to 16 total epochs
Closing dataloader mosaic
Image sizes 1024 train, 1024 val
Using 16 dataloader workers
Logging results to /home/patrick/ml/bishop/runs/y12_phase4_mixed20k_ms102
Starting training for 16 epochs...

        Epoch     GPU_mem   box_loss   cls_loss   dfl_loss  Instances        Size
        10/16       0.545G      1.619      1.112      1.324          5        1024: 100% ---------- 3334/3334 3.1it/s 17:53<0.4s
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100% ---------- 34/34 3.8it/s 8.9s0.2s
               all        400        995      0.829      0.701      0.773      0.401

        Epoch     GPU_mem   box_loss   cls_loss   dfl_loss  Instances        Size
        11/16       12.6G      1.596      1.067      1.312          1        1024: 100% ---------- 3334/3334 3.4it/s 16:19<0.7s
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100% ---------- 34/34 3.8it/s 9.0s0.2s
               all        400        995      0.827      0.707      0.776      0.405

        Epoch     GPU_mem   box_loss   cls_loss   dfl_loss  Instances        Size
        12/16       14.3G      1.583      1.038      1.302          6        1024: 100% ---------- 3334/3334 2.0it/s 27:17<0.7s
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100% ---------- 34/34 3.4it/s 10.1s.2ss
               all        400        995      0.841      0.709      0.783      0.409

        Epoch     GPU_mem   box_loss   cls_loss   dfl_loss  Instances        Size
        13/16       13.9G       1.56      1.004      1.294          7        1024: 100% ---------- 3334/3334 0.7it/s 1:17:23<1.9ss
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100% ---------- 34/34 2.6it/s 13.2s0.2s
               all        400        995      0.838      0.713      0.788      0.413

        Epoch     GPU_mem   box_loss   cls_loss   dfl_loss  Instances        Size
        14/16       14.3G      1.537     0.9718      1.278          3        1024: 100% ---------- 3334/3334 1.7it/s 32:46<0.8ss
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100% ---------- 34/34 3.7it/s 9.1s0.2s
               all        400        995      0.839      0.717       0.79      0.415

        Epoch     GPU_mem   box_loss   cls_loss   dfl_loss  Instances        Size
        15/16       13.9G      1.511     0.9412      1.264         15        1024: 100% ---------- 3334/3334 0.9it/s 1:03:16<1.4ss
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100% ---------- 34/34 4.8it/s 7.1s0.2s
               all        400        995      0.847      0.718      0.795      0.418

        Epoch     GPU_mem   box_loss   cls_loss   dfl_loss  Instances        Size
        16/16       13.9G      1.509     0.9307      1.264         60        1024: 100% ---------- 3334/3334 0.7it/s 1:18:26<1.9ss
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100% ---------- 34/34 2.9it/s 11.8s0.2s
               all        400        995       0.86      0.709        0.8      0.421

7 epochs completed in 5.243 hours.

Iteration 2

Week 4 - Detect, Track & Visualize

The design needed to prioritize scalability, efficiency, and clear separation of concerns, so that video ingest, inference, and visualization could scale independently as usage increased.

Inference video frames with the trained model

Track multiple detections over time

V4 🪄

Detect, Aggregate, Coalesce -> Track

Visualize human tracks on the client video

Iteration 3

Week 6 • Create, View, Manage & Associate

Create projects and associate them with videos

Bishop project creation, upload, and project management screens

View AI detections in the video, synced to the timeline

Manage client video library

Bishop video library management and delete confirmation screens

Associate detections with the SAR mission map

The Minimal, but Viable Product

Key Learnings

Speeds beats Perfection

Operators valued fast, "good enough" detections over slower, marginally more accurate ones

Transparency builds trust

Showing ML confidence scores helped operators calibrate trust and make better decisions

Resilience in the field

Offline capability and a simple UI is critical for use in remote environments

Thank you

To the SAR operators who trusted us with their feedback, field-tested early prototypes in difficult conditions, and never stopped pushing us to build something that saves lives!!

To my Co-Founder and Lead Designer Shelby Reilly who slayed until the very end! :)

And to the teams who keep moving when the clock doesn't stop.

Next case study

01 AEROT

Role: Senior Software Engineer
Timeline: 2024 - 2026
Stack: Full-stack, C++, Typescript, React

Friendly force EM training for the United States Marine Corps.

View details