Basketball AI · Gallery

NetsReview v2.4 · session live

Gallery

0Photos

0Tags

0Pending

[ no photo selected — tag one from Lightroom to populate this view ]

File—

Filename—

Captured—

Status—

JerseyYOLO + OCR

—

Roster · Brooklyn Nets

—

YOLOPose · Ball · Hoop · Players

No YOLO data for this photo.

CLIPPrescreening scores

No CLIP scores for this photo.

ClaudeAI reasoning

◆ AI read

Tag a photo from the Lightroom plugin to see Claude's reasoning here.

Applied0 tags

Pendingheld · low conf

Add missed tagmanual

Mutual Exclusions

Pairs the AI should never apply together.

0 rules

←→

Confidence Thresholds

Minimum level required to auto-apply each tag.

0 entries

Override for

Required Combinations

If a tag applies, automatically also apply another tag.

0 rules

→ also apply

Plugin Activity

Idle

0 calls

last: —

Time Filename Result Dur

No activity yet — run "Tag Selected Photos with AI" from Lightroom

01 The Lightroom Plugin

What the plugin does

The plugin lives inside Adobe Lightroom Classic. It automates keyword tagging of sports photos using Claude AI — with human feedback that makes it smarter over time.

AI tagging in one click — Select photos, hit Tag with AI. Each photo is sent to Claude with game context (opponent, jersey edition, roster) and gets keyword tags back.

Writes to LR catalog — Tags become native Lightroom keywords, instantly visible in the keyword panel and ready to filter and export.

Give Feedback dialog — Review each AI decision: ↑ Keep, ✗ Remove, or ＋ Add a missed tag. Decisions write to LR and train the gallery on the server.

Sync from Web — Reviews done on photoai.work (any device) are pulled into Lightroom with a single menu click.

CLIP ViT-B/32 ONNX claude-sonnet-4-6 6 positive examples / tag 4 negative examples / tag ~1 API call / photo

Lightroom Classic — Library Menu

02 photoai.work

The web platform

The cloud layer that makes the plugin smarter over time — gallery training, web review queue, and R2-backed persistence that survives server restarts.

🖼

Gallery Training

Every feedback decision adds the photo as a positive or negative example for that tag. The AI sees these examples on the next tagging run — it improves with every session.

🔍

Web Review Queue

Tagged photos appear here for review on any device. Approve or reject on a phone courtside — decisions queue server-side and sync to LR on next pull.

☁

R2 Persistence

Gallery examples, tag definitions, rules, keyword-edit queues, and feedback decisions are all backed up to Cloudflare R2 — Railway can redeploy without losing state.

03 Integration Flow

How it all connects

Two primary workflows. Branch A is the shoot-to-review loop inside Lightroom. Branch B covers every sync path between the web platform and the LR catalog.

NetsAI Lightroom Plugin

BasketballTagger.lrplugin · photoai.work · Railway · Cloudflare R2

A — Shoot & Tag → Review

Tag Photos with AI · Give Feedback

Select photos in LR grid → "Tag Selected Photos with AI"

Export JPEG · base64-encode · 2 concurrent workers

CLIP ViT-B/32 Pre-screener · No API call

Runs server-side via ONNX — scores all 42 tags by cosine similarity between the photo embedding and gallery examples. Tags with ≥3 gallery examples use visual similarity; tags with fewer fall back to text description embeddings.

Phase 1 — top-5 by normalized score. Phase 2 — add best tag from any unrepresented semantic category (score ≥ 0.40 floor), ensuring each visual group has a voice.

42 tags scored 5–11 candidates gallery path · text path 6 semantic categories

claude-sonnet-4-6 Single call · Full gallery

One call per photo — Sonnet sees the full gallery (6 pos + 4 neg) for every candidate, plus tag definitions, jersey context, roster, and rules. Replaces the old 5–8 call bucket pipeline entirely.

Output: { reasoning, keywords, confidence }

6 positive ex / candidate 4 negative ex / candidate 1 call / photo

rules.json · Rules Engine

Confidence thresholds filter low-certainty tags. Required combinations fire (e.g. Driving → Look). Mutual exclusions enforced — higher-confidence tag wins.

claude-haiku-4-5 Conditional · Mutex re-query

Only fires when the rules engine detects a mutual exclusion conflict (e.g. dunk + block on the same photo). Focused 2-tag call with no mutual exclusion rules — lets Haiku adjudicate the specific contested play without interference.

fires ~10% of photos 2 tags · focused context

POST /api/tag-photo · /api/session-upload → { dataUrl, filename, gameContext } ← { keywords, pendingReview, reasoning }

✓ Confirmed keywords written to LR catalog · checkpoint saved

Give Feedback

Give Feedback… in LR menu · card UI per photo (up to 50)

↑ Keep ✗ Remove ＋ Add missed

POST async /api/feedback ↑ { tag, polarity:"pos" } ✗ { tag, polarity:"neg" } ＋ { tag, type:"missed" } ← photo added to gallery training examples · saved to R2

✓ Single withWriteAccessDo after Done · Keywords applied to LR · Gallery updated

B — Web Sync

Three paths · same contract: keep LR ↔ photoai.work in sync

↓ Pull Down
Sync from Web

Sync from Web in LR menu

GET /api/keyword-edits
?consume=1 ← { edits: [{ filename,
tag, action }] } atomically dequeues

R2-backed queue
Tags.normalize()
createKeyword()

✓ Keywords added /
removed in LR

↑ Push Up
Web Review → LR

Thumbs-up/down on
photoai.work/status
any device

POST /api/keyword-edits → { filename, tag,
action: add|remove } ← queued in R2

_keyword_edits
+ _feedback_decisions
survives restarts

Run ↓ Pull Down
in LR to apply

✓ Keywords applied
to LR catalog

⏪ Backfill
All-time history

Sync Reviewed → LR
button on photoai.work

POST /api/sync-approved ← { queued: N }

Reads all-time
_feedback_decisions
→ rebuilds queue
deduplicates per
filename + tag

Run ↓ Pull Down
in LR to apply
all N edits

✓ All historical
reviews applied
to LR catalog

Jersey Edition Labeler

Click a photo · assign a class · repeat

0 photos labeled