60 Photos and videos
started making a small list of people on here who are actually building things and worth following
every week i’m going to share the best small builds i find
if you want to be in next week’s round, reply with a screenshot
x.com/i/lists/20655157556058…
#buildinpublic
2
3
210
leon retweeted
Jun 12
wow sick new billboards from openai
95
251
5,122
2,823,736
started making a small list of people on here who are actually building things and worth following
every week i’m going to share the best small builds i find
if you want to be in next week’s round, reply with a screenshot
x.com/i/lists/20655157556058…
#buildinpublic
2
3
210
and last there’s Corey, @TechTalentBP, building Micro Coach
an ai fitness app for training plans, workout tracking, and actually seeing your progress
crowded space, but very real problem. most people do not need more fitness content, they need help knowing what to do next
1
1
1
110
if you have no dataset yet, start with text anchors.
write prompts for what your app cares about:
“a photo of a damaged package”
“a close-up photo of powdery mildew”
“a receipt on a table”
“a scratched trading card”
embed those prompts, embed your test images, compare by similarity.
that gives you a fast first read. but once a category matters, collect 5-10 real reference images per label. use actual user-style photos: weird angles, bad lighting, clutter, partial views.
perfect examples make fake demos. messy examples make useful apps.
1
72
the biggest jump in my Stampede! experiments came from replacing text-only anchors with image-reference centroids.
text anchors proved the idea, but only ~50% of holdouts were safe enough to accept. after switching to image-reference centroids, safe accepts jumped to 95% and false accepts dropped to 0.
the workflow i’d steal:
test encoder offline, start with text anchors, add real reference images, keep separate holdouts, add near-misses that should fail, and only accept when the top match clearly beats the runner-up.
follow along if you want more notes from my journey learning how to train, ship, and actually use open-source models in real apps.
55
if you’re building image recognition into an iOS app, here’s one result i didn't expect:
let's dive into some of my research experiments with Stampede!
i assumed the best model input would be a clean cutout of the animal from Core ML
it turned out to be one of the worst
i tested this while building local animal recognition for Stampede! using BioCLIP 2 in Core ML, which is Apple’s framework for running models locally on iPhone
1
88
the cleaner looking cutout input was worse
the model didn't want a perfect sticker silhouette. it wanted the animal plus some surrounding context
that lines up with how CLIP-style models are trained. background and scene cues are often part of the representation, so removing too much context can make the model less reliable, not more
1
45
the practical takeaway was to separate UI from ML
for the UI, the cutout is great. it looks like a collectible sticker
for the model, i now use a wide saliency crop with 65% padding
also, every preprocessing change changed the score distribution, so retuning thresholds after each crop variant mattered almost as much as the crop itself
if you’re building your own image recognition flow on iOS, i’d test crop strategy much earlier than you think
follow if you liked this story, and want more research notes on my journey to learn how to train custom models
40
the intuition seemed obvious
if the app can isolate the animal, remove the background, and feed the model just the subject, that should make recognition easier
so i expected the ranking to be:
1. exact animal cutout
2. tight crop around the animal
3. full photo
the benchmark said otherwise
10