Sprint 3 Week 1: Pipeline Integration, Report Planning, and Viva Preparation
As we wrote in last week’s retrospective, Sprint 3 would open with RAG pipeline integration — and that’s exactly what filled this first week. Alongside it, the Data and Model Layer teams pushed key groundwork forward, we locked in two roadmap decisions, and we started planning ahead for the viva and the group report.
Sprint 3 Week 1 Progress
Data Layer. Lei and Qiuting completed feature engineering, baseline statistics, and proxy signature replay analysis — the groundwork the Model Layer needs to keep improving anomaly detection.
Model Layer. Ray and Lucca are finalising the layer interface, verifying data quality, and updating the ingestion pipeline, ahead of the new prediction fields described below.
Report Layer and Dashboard. This was the big push of the week. Charlotte and Jintong completed full RAG pipeline integration: a ChromaDB knowledge indexer covering 28 documents across all seven anomaly types, metadata-filtered retrieval — the approach we chose over semantic search back in Sprint 2 — confidence-based language guidance, signal correlation analysis, few-shot examples, an automated report quality evaluator, and a prompt chain validator.
With the pipeline built, we ran our first RAG-versus-baseline evaluation across three test scenarios and saw RAG-enhanced reports consistently outscore the baseline, especially on using appropriately cautious language in ambiguous cases. A full breakdown of the retrieval evaluation methodology and results is in the next post.
Key Decisions This Week
Working with the Model Layer, we confirmed two changes to what the system will predict.
New output fields. risk_score, failure_probability, and estimated_cycles_to_failure will now all be produced, and each answers a different question:
| Field | What It Tells You |
|---|---|
| risk_score | How abnormal things look right now |
| failure_probability | How likely a failure is coming |
| estimated_cycles_to_failure | Roughly how much time is left |
Three distinct outputs the Model Layer will produce, each answering a different question.
Both new fields are required by the IBM project brief, and unlike risk_score, they depend on forecasting future behaviour rather than scoring the present moment.
risk_score history moves to the Model Layer. History was originally going to live in the Dashboard, but the Model Layer uses IBM Granite TTM — a time-series foundation model — to predict how risk will trend over time. That’s far more accurate than the Dashboard doing simple linear extrapolation from stored history, so the Model Layer now owns it.
Viva Preparation
On our supervisor’s advice, we’re starting viva preparation now rather than later — as soon as there’s a usable product to show, not after everything is finished. The plan is a short video demonstration of the dashboard in action, alongside slides covering the technical details underneath. Nothing needs to be polished yet; the goal is to start early and improve gradually as the system matures.
Group Report Planning
We’ve also agreed on chapter ownership across all six of us. To keep six people writing smoothly together, we’re using Overleaf — through Bristol’s institutional Premium licence — for real-time collaboration, GitHub for version control, and a shared Zotero group library for references, all following the Bristol memoir report template.
What’s Next
With the RAG pipeline live and validated, Sprint 3 continues with expanding dashboard coverage to all seven anomaly types, wiring in the new prediction fields, and steadily building out the viva materials and group report alongside the technical work.