AICHE +

C

Comet Integration

The short answer: open Comet's experiment view, click into the Notes tab or any text field, press ⌃+⌥+R (Mac) or Ctrl+Alt+R (Windows/Linux), speak your observations for 30-60 seconds, and AICHE inserts formatted documentation of your experiment.

ML engineers run experiments in batches. You kick off a training run, adjust hyperparameters, observe the metrics, and start another. The cycle is fast, but the documentation lags behind. Every experiment deserves a note explaining the hypothesis, why you chose those parameters, what you observed, and what to try next. This documentation is critical for reproducibility and for your future self who'll revisit these experiments in three months. But typing detailed notes during a training run is awkward. You're switching between terminal windows, Comet dashboards, and notebooks. Voice captures observations without interrupting the work.

1. Open Comet ML in your browser and navigate to your project or experiment view.
2. Click on an experiment to open its detail page.
3. Click into the Notes tab, or click the text field in any section that accepts notes.
4. Press your AICHE hotkey (⌃+⌥+R on Mac, Ctrl+Alt+R on Windows/Linux) to start recording.
5. Speak your observations: what you're testing, why you chose these parameters, what the early metrics show, and what you plan to try next.
6. Press the hotkey again. AICHE transcribes your speech and inserts it into the notes field.
7. Continue monitoring your experiment. Add additional notes as training progresses by clicking into the field and dictating again.

Experiment Hypotheses and Parameter Rationale

Every experiment starts with a hypothesis, even if it's informal. "I think increasing the learning rate to 3e-4 will reduce training time without hurting validation accuracy because the loss curve shows slow convergence at 1e-4." Most ML engineers hold this reasoning in their head and never write it down. When the experiment finishes and you're comparing results in Comet's comparison view, the numbers say what happened but not why you tried it. Voice makes logging the hypothesis trivial. Before you start a run, open the experiment in Comet, press your hotkey, and speak: "Testing learning rate 3e-4, up from 1e-4. The previous run showed the loss plateauing after epoch 5 with minimal improvement through epoch 20, suggesting the learning rate is too conservative. Keeping batch size at 32 and the same augmentation pipeline. Expecting faster convergence and similar final accuracy. If validation loss starts increasing before epoch 10, the rate is too aggressive and I'll try 2e-4 as a middle ground." That's 20 seconds. Now when you compare this experiment against others in Comet, the context is right there.

Real-time Training Observations

The most valuable experiment notes are the ones captured during training, not after. You notice something in the loss curve, an unexpected spike at epoch 12, validation accuracy plateauing while training accuracy keeps climbing, a sudden drop in a custom metric. These observations are time-sensitive. If you wait until the run finishes to write them down, you've lost the specificity: which epoch, which metric, what the dashboard looked like at that moment. Voice lets you document in real time. You see the anomaly, press your hotkey, and speak: "Epoch 12 shows a loss spike to 2.3 from a baseline of 0.8. Validation accuracy dropped 4% simultaneously. This happened around the same epoch in run 47 when I used the same data augmentation. Suspecting the augmentation pipeline is introducing corrupted samples at this batch range. Will inspect the training data for epochs 11-13 after this run completes." That note, timestamped alongside the experiment metrics in Comet, becomes the kind of documentation that makes experiments reproducible.

Model Comparison Notes and Conclusions

After running a batch of experiments, Comet's comparison view lines up metrics side by side. The numbers tell part of the story, but the conclusions need to be written. Which model won and why? What should the next round focus on? Typing these comparison notes takes 10-15 minutes, which feels like overhead when you're eager to start the next batch. Voice makes writing conclusions fast. Open the comparison view, look at the metrics, press your hotkey, and speak: "Run 52 with the transformer architecture outperforms the CNN baseline by 3.2% on validation accuracy but takes 4x longer to train. For production deployment where inference latency matters, run 48's distilled model is the better choice. It's only 1.1% behind on accuracy but runs at 8ms inference versus 45ms. Next batch should explore quantization on the distilled model to see if we can get inference under 5ms without dropping below the accuracy threshold." That's a 20-second dictation that saves hours of re-running experiments to reconstruct the decision.

Heads-up: Comet's Notes field accepts plain text. AICHE inserts your transcribed speech directly. If you want structured notes, enable Content Organization in AICHE settings to get automatic paragraph breaks and section-like formatting.

Pro tip: dictate notes at three specific moments: before the run (hypothesis and parameters), during the run (observations and anomalies), and after the run (conclusions and next steps). This three-point pattern creates a complete experiment narrative that makes your Comet project a genuine research log.

Result: experiment documentation happens at the speed of observation instead of typing. Your Comet project becomes a searchable record of not just what you ran but why you ran it, making experiments reproducible months later.

Do this now: open Comet, navigate to your most recent experiment, click into the Notes tab, press your hotkey, and dictate what you tested, what you observed, and what you'd try next.