A disabled button was not a broken recorder.
The Mac app was telling the truth about paired capture while hiding the working single-camera path below the preview.
6 minRead
Find your next useful idea.
Field notes, research, playbooks, and lessons. One place to follow the threads that matter to your work.
The copy had to become permission to delete.
Moving a golf video off the iPhone was easy. Proving that the phone copy could be safely removed was the real system.
7 minRead
Seventy-six UI tests passed. Two range controls were still outside the window.
Swing Trainer's Live Studio passed its presentation suite, but a real 1149×772 window clipped critical capture controls. Why visual verification remains a separate release gate.
5 minRead
I marked a swing, closed the app, and discovered the annotation was never a record.
Swing Trainer needed to preserve partial human review without calling incomplete work training truth. The revision model that keeps drafts, labels, and source video separate.
8 minRead
I taught the app when to record a swing. I refused to let it call that motion Takeaway.
Body motion can start and stop a golf recording, but it cannot prove where the club is. The evidence boundary behind Swing Trainer's hands-free capture system.
7 minRead
A phone can trace the club. That does not mean it measured clubhead speed.
The physics boundary inside Swing Trainer: what monocular video can support, what calibration might unlock, and what still requires an external instrument.
8 minRead
I had enough data to spend money on a GPU. I did not have enough evidence to train.
Why the Swing Trainer RunPod gate stops at provenance, frozen splits, and real five-point labels—even with thousands of images already staged.
7 minRead
We found 8,295 golf images. Almost none were the data we needed.
GitHub, Roboflow, Kaggle, and Hugging Face produced plenty of files. The hard part was refusing to call mismatched labels a dataset.
8 minRead
I thought tracking a golf club meant finding a line. Then the line disappeared.
The club looked simple until motion blur, hands, clothing, and the edge of the frame turned one object into five different evidence problems.
7 minRead
Do not put the club in the body frame.
A pose system can see a body, a club detector can see a club, and a launch monitor can report a ball. Calling all three one measurement is how a golf product starts inventing facts.
5 minRead
I called three sign-recognition models failures. The recipe was the failure.
A Parley notebook reported three landmark architectures as broken on cross-signer ASL. A warmup and a gradient clip brought all three back, and two matched the best model. The ranking had measured my training recipe, not the models.
7 minRead
The 38-point gap: one accuracy number, twenty-one very different users
Our sign model averages 42% across signers. That average hides a range from 26% to 64% — and the thing that decides where a person lands is not the signs they make, it is who they are.
7 minRead
45%, not 90%: the only sign-recognition number I trust
Our best landmark-only sign model scores 45% on signers it has never seen. The field routinely reports numbers twice that high. The lower number is the honest one, and it is the one we publish.
6 minRead
Six ways hearing-built sign-language AI fails the Deaf community
I keep a running catalog of how hearing-led sign-language AI fails. It is not a list of other people's sins. It exists so Parley can catch itself the moment it starts to look like one of them.
7 minRead
Training a custom CV model with Claude: the data quality lesson we learned the hard way.
Clean data beats model size. Every time. Don't upgrade the model until you've audited the labels.
7 minRead
Claude built our CV pipeline. Then it lied about being done.
Agents are unreliable judges of their own work. Here's how a structural fix — not a smarter model — stopped QC's CV pipeline from shipping silent failures.
8 minRead
Computer vision already runs elite sports. It's about to run the rec league too.
Twelve-camera tracking rigs and Hawk-Eye are infrastructure at the top. The same capability now fits on a $249 board and a commodity camera. What that unlocks across every sport — and why the smartest way in is the narrowest one.
7 minRead
From Pixels to Club Physics: An Evidence-Bounded Architecture for Monocular Golf-Club Pose, Kinematics, and Training Data
What a phone camera can tell us about a golf club—and where physical measurement needs more evidence.
18 min readRead
Recipe, Not Architecture: Three "Failed" Sign-Recognition Models Recover Under Per-Architecture Training
Three sign-recognition models recovered when the training recipe changed. A closer look at what model comparisons actually measure.
12 min readRead
The Honest Ceiling on Landmark-Only Isolated-Sign Recognition: A Signer-Holdout Audit
An audit of sign recognition on unseen signers, with multiple seeds and explicit limits on what the results establish.
13 min readRead
Signer-Dialect Robustness in Landmark-Only Sign Recognition: A 21-Fold Leave-One-Signer-Out Study
Why one average accuracy score can conceal very different experiences across 21 signers.
14 min readRead
Real-Time Computer Vision for Bounded-Court Scoring: Architecture, Tradeoffs, and Production Lessons
Real-time scoring of physical-object sports requires a computer vision pipeline that can simultaneously satisfy constraints from three domains: inference latency (sub-35ms end-to-end), detection reliability (>99% on custom domain data), and commercial licensing (Apache 2.0 for network-service deployments).
14 min readRead
Data-Centric CV Pipeline Engineering: Lessons from 15 Training Iterations on a Custom 3-Class Detection Domain
Generic computer vision model fine-tuning treats dataset construction as secondary to architecture selection.
15 min readRead
Same Task, Three Harnesses: A Throughput-Quality-Cost Benchmark of Claude Code, Codex, and Cursor on a Multi-Stage CV Pipeline Task
We report a case study (n=1 task, three agent harnesses, single operator) comparing Claude Code, OpenAI Codex CLI, and Cursor on the construction of a multi-stage computer-vision evaluation pipeline.
16 min readRead
Detecting Eval-Train Overlap in Production CV Pipelines: A Lightweight Audit Protocol
Computer vision systems shipped to production frequently report misleading evaluation metrics due to undetected overlap between training and held-out eval data.
14 min readRead
Multi-Stage Fallback in Real-Time Computer Vision: A Methodology Study
Real-time computer vision systems that must operate under latency constraints face a tension between detection accuracy and inference speed.
14 min readRead
The CV-for-Physical-Sports Wedge Playbook
Don't build computer vision for sports. Build it for one sport — the one with the most constrained geometry and the simplest rules — get it referee-grade, then generalize. Three filters and the trust-bar framework.
DisciplineRead
The Honest Sign-Language Research Playbook
The discipline behind Parley's notebooks: question-first contracts, signer-holdout splits, multi-seed floors, failure-modes-first, and a Deaf-community honesty checklist. The rules that make a 45% you can trust beat an 85% you can't.
DisciplineRead
Picking a computer-vision wedge: start with a fixed playfield
The hardest version of a CV problem is an open field; the easiest is fixed geometry you can pin with homography. How to choose the narrow, solvable entry point — then generalize.
11 min read · 20 min applyRead