Machine Learning Use Cases (ML)¶
UC-ML-001a: Curate Training Dataset¶
Purpose: Select and prepare high-quality data for fine-tuning.
| Property | Value |
|---|---|
| Actor | Data Engineer |
| Trigger | Weekly Schedule |
| Priority | P2 |
Main Success Scenario:
1. Query "Golden Dataset" (UC-ML-003) for approved corrections
2. Filter for audio quality (SNR > 20dB) and transcript length
3. Split into Train/Val/Test sets (80/10/10)
4. Convert to HuggingFace Dataset format
5. Version control the dataset artifact (DVC)
Acceptance Criteria: 1. [ ] Zero PII leakage in training set 2. [ ] Balanced distribution of dialects
UC-ML-001b: Execute Fine-tuning Run¶
Purpose: Run the compute-intensive training job.
| Property | Value |
|---|---|
| Actor | ML Ops Pipeline |
| Trigger | Dataset Versioned |
| Priority | P2 |
Main Success Scenario:
1. Provision GPU cluster (e.g., A100s)
2. Load base model (Whisper/Llama) and new dataset
3. Execute LoRA/QLoRA training loop
4. Log metrics (Loss, WER) to MLflow
5. Save model checkpoints to Model Registry
Acceptance Criteria: 1. [ ] Auto-shutdown of GPUs after completion 2. [ ] Alert on gradient explosion or NaN loss
UC-ML-002: Dialect Evaluation & Benchmarking¶
Purpose: Measure model performance across specific Indian languages/dialects.
| Property | Value |
|---|---|
| Actor | QA / ML Team |
| Trigger | New Model Candidate |
| Priority | P1 |
Main Success Scenario:
1. Load benchmark datasets (Hindi, Telugu, Tamil, etc.)
2. Run inference with candidate model
3. Compute WER, CER (Character Error Rate) per dialect
4. Generate report comparing vs baseline
5. Flag regressions > 1%
Acceptance Criteria: 1. [ ] Covers top 5 target languages 2. [ ] Includes medical-specific vocabulary test 3. [ ] Automated pass/fail gates
UC-ML-003: Continuous Quality Feedback Loop¶
Purpose: Improve models using doctor corrections.
| Property | Value |
|---|---|
| Actor | System |
| Trigger | Doctor edits note |
| Priority | P2 |
Main Success Scenario:
1. Capture "Diff" between generated note and final signed note
2. Anonymize and store as (Input, Correction) pair
3. Aggregate corrections by category (Hallucination, Missed Entity)
4. Add high-quality pairs to "Golden Dataset"
5. Trigger fine-tuning (UC-ML-001) when dataset grows by 10%
Acceptance Criteria: 1. [ ] Strict PII scrubbing before adding to training set 2. [ ] Doctor opt-in/opt-out for data usage 3. [ ] Quality filter to exclude bad edits