Transforming 30+ Computer Vision Algorithms into a Curator-Controlled Workflow

Professional photographers managing high-volume event and sports shoots spent up to 18.5 hours per job manually reviewing and selecting images. Machine learning could score and filter that work automatically—but earlier automated approaches had already failed, because photographers wouldn’t trust selections they couldn’t interrogate or correct.

The platform’s intended advantage was its more than 30 computer-vision and machine-learning algorithms. In practice, that capability was the design problem. Each algorithm produced its own opinion about what mattered—quality, faces, duplicates, themes, priority—and none of it meant anything to a photographer who simply needed to decide which images to keep. More algorithms didn’t make the product more usable; they made its decisions harder to trust.

The central design decision I drove was to stop treating this as an automation problem. Photographers didn’t need access to 30 algorithms, or a system that decided for them. They needed a small set of understandable controls tied to decisions they already made: what to include or exclude, how strict quality should be, how images should be grouped, and which of the system’s picks deserved a second look. I led UX strategy and product design direction—where AI should assist rather than replace, how recommendations were surfaced, and the principle that the photographer’s judgment stayed authoritative.

Outcome — Reduced reported review time from an average of approximately 18.5 hours to about 1 hour — based on photographers’ actual-use estimates — while preserving photographer control.

Absorbing the complexity, not exposing it

Behind the scenes, dozens of algorithms each produced their own interpretation of what mattered:

• Image Quality Analysis
• Face Detection & Recognition
• Near-Duplicate Detection
• Event & Theme Clustering
• Image Prioritization

One workflow had to turn all of it into fast, confident decisions. The tradeoff was whether to expose that machinery—giving photographers direct access to algorithm outputs and parameters—or absorb it, translating dozens of signals into a handful of controls tied to decisions photographers already made. Exposing the algorithms would have showcased the technology; it would also have made the product unusable. I chose to hide the pipeline and design to the decision: photographers would control outcomes, not the 30 algorithms underneath them.

The design challenge became one of orchestration: translating complex machine outputs into a workflow professionals could understand and trust.

Curator-Controlled Workflow and Experience Architecture
The orchestration layer I defined absorbs 30+ algorithm outputs into a single calibration step, then hands the photographer a prioritized workflow they steer and export from—the structural decision behind the drop from approximately 18.5 hours to about 1 hour.

Useful defaults, expert calibration

A tool that forced photographers to configure 30 algorithms before their first job would never get used; one that decided everything for them would never be trusted. The import step resolves that tension. Every job starts from working defaults—cull rate, grouping, duplicate handling—that a photographer can run untouched, or open up and recalibrate control by control. Defaults make the product learnable on day one; the same controls let an expert tune it to a specific shoot.

Photographers could:

• Accept sensible defaults or override any of them
• Adjust how aggressively images were culled
• Choose how photos were grouped
• Set the minimum and maximum size of each group
• Recalibrate without restarting the workflow

Full automation was technically achievable, but prior attempts showed photographers wouldn’t rely on outputs they couldn’t audit or override. The bet I made was that calibrated control, not raw speed, was what would make the automation usable.

The import screen starts from working defaults—cull rate, grouping, duplicate handling—that a photographer can accept as-is or recalibrate. Defaults make it usable immediately; the controls behind them keep expert judgment in charge.

Making the system’s thresholds visible

The system scores every image and sorts thousands of candidates into review sets. But a score is the algorithm’s estimate of quality, not a verdict on it—a technically sharp frame can still be the wrong one to keep. So the decision was to expose that estimate rather than act on it silently: photographers see the score range behind each group and can widen or tighten it, which makes the line between “good enough” and “not” theirs to set, not the model’s. Speed comes from the sort; confidence comes from being able to see the threshold and move it.

A score-range slider sets which of the system’s quality estimates land in each group and updates the set immediately. The threshold is the photographer’s to move—the score informs the decision rather than making it.

Recommendations, not decisions

Near-duplicate detection is where automation is both most tempting and most consequential: the system is usually right about which frames are near-identical, and occasionally wrong about which one is worth keeping. It would have been simple to auto-delete the extras. Instead the system groups the duplicates, marks its recommended pick, and leaves both the choice and the reversal to the photographer. That restraint is the product-design decision—on the call where the model’s confidence and the stakes are both highest, the human stays in control rather than the automation.

In a duplicate group, the system marks its recommended selection but keeps the alternatives one click away. The recommendation is visible and reversible; the photographer makes the final call.

Impact

One outcome is based on photographers’ actual-use estimates; the rest describe what the workflow was designed to do.

• Reported: reduced reported review time from an average of approximately 18.5 hours to about 1 hour, based on actual-use estimates provided by several photographers, with the photographer retaining control over final selection
• Delivered: dozens of independent machine-learning outputs were unified into a single, coherent curation workflow
• Designed for: removing repetitive review effort while keeping creative judgment with the photographer
• Intended: turning per-job review from days into roughly an hour was meant to support the operational scale of high-volume photography businesses—a design goal, not a measured business result

My Leadership Scope

• Led UX strategy and product design direction for an AI-assisted curation platform, defining the experience architecture across a multi-algorithm ML system
• Proposed the curator-controlled interaction model and the human-AI trust approach: the calibration, confidence signaling, and override patterns that kept photographers in control of final decisions
• Directed the end-to-end product-design effort—interaction design and prototyping—working hands-on to resolve high-stakes decisions at the human-AI boundary
• Shaped validation: prototype sessions with working photographers tested whether calibrated control resolved the trust barriers that had caused earlier automated approaches to fail
• Partnered across product, engineering, data science, and business stakeholders on what the system would and wouldn’t do—translating competing ML capabilities into a product experience photographers could trust
• Helped determine where automation, defaults, calibration, and manual review each belonged, informed by research on where algorithmic intervention accelerated the workflow versus introduced friction or eroded trust