AMIRA

The problem

Fractionation process engineers needed near-real-time insight across FP002, FP005 and G-Cent standard work, but the data lived in disconnected systems (OSI PI sensors, JDE commodity data, LIMS) and answering a single process question meant manual pulls and spreadsheet wrangling. PowerBI dashboards showed what happened but couldn't explain why — or be trusted to quote exact numbers.

Architecture

Reliability and scale

• Refreshed on demand in near real time; every pipeline runs on a three-hour schedule, and data-freshness badges are per session
• AWS S3 to Databricks Unity Catalog migration behind a storage abstraction with a schema manifest that preserves R types across SQL round-trips; no downstream module changed
• 54 modules and roughly 84,000 lines of production R, with a centralized RAG helper module so retrieval logic is implemented once
• Self-hosted fonts and a responsive viewport because plant-floor networks block CDNs
• Approved through the site SDLC as a non-GxP informational tool; the 16-artifact dossier closed a platform audit finding

What it does

AMIRA — Advanced Manufacturing Insights & Real-time Analytics — is a multi-process analytics platform covering FP002, FP005 and G-Cent standard work, refreshed on demand in near real time with every pipeline on a three-hour schedule: 54 modules and roughly 84,000 lines of production R, built and maintained end to end. A Tier 2 Deep Dive dashboard identifies constraints with SPC control limits, XmR charts, Pareto views and cross-filtering; constraint analytics on the cleaning step supported a ~30% improvement at a bottleneck constraint, returning engineer and operator hours to the process. Alongside it sit standard-work analytics per process, delay reporting written back to Databricks, JDE commodity reconciliation against process data, and two assistants — TOMO for text and TOMO Vision for charts.

TOMO — the AI assistant

TOMO is a hybrid-RAG chatbot built on a discipline I use everywhere: R computes the statistics (mean, median, percentiles, trend, shift) and the LLM only interprets them. That single decision cuts token spend ~70% and gives a hard anti-hallucination guarantee — the model cannot invent a number it was never handed. TOMO pre-filters by equipment before the vector search and applies date ranges client-side afterward.

• TOMO Vision — multimodal reasoning over charts, tables and text together, with chart-aware MVDA interpretation
• Voice input (Web Speech API) for gloved operators; TTS that skips table rows so operators hear analysis, not recitations
• Thumbs-up/down feedback and last-5 recent prompts logged to Databricks Delta for one-click reuse

Process modeling

• Sensor Process Variation — OSI PI minute-resolution traces (flow rate, differential pressure, temperature, turbidity) with P15–P85 'Good' envelopes and per-lot anomaly overlays
• SIMCA-style Batch Evolution Model — DModX-over-maturity trajectories against the Good-population +3SD limit, with per-minute sensor contribution drill-downs (Hotelling's T², Q-residuals)
• Multi-lot drill-down — overlay up to 5 lots; the LLM payload is capped at 3 to bound token spend

Engineering & accessibility

A centralized RAG helper module consolidates 12 core functions — vector search, post-filtering, date parsing, R-powered summary stats and scheduling-delay enrichment — so the same logic isn't reimplemented across features. The UI ships a viridis colorblind-safe high-contrast mode (WCAG 2.1 AA), 44px+ tap targets, and data-freshness badges shared across users.

2026: models, migration, governance

• MVDA classifier + SHAP answers 'why was this lot long?' — a three-engine shootout (Random Forest, XGBoost, penalized logistic regression) with held-out metrics, a per-lot SHAP waterfall, and performance-over-time fed by an append-only prediction log
• AWS S3 → Databricks Unity Catalog migration behind a storage abstraction with a schema manifest that preserves R types across SQL round-trips — no downstream module changed
• UX 2.0: self-hosted fonts because plant-floor networks block CDNs, responsive viewport, phase-colored role cards, KPI tiles
• August 2026: a dedicated SPC engine shipped with its own tests — 90 tests across six testthat files, including a determinism regression asserting TOMO's R-computed blocks against seeded fixtures
• October 2026, v1.10.2 'Clarity & Accessibility Enhancements': KPI tiles with reference and n, honest per-session freshness (a shared timestamp had made one user's refresh look current in every other session), SPC rules 1–4 with labelled limits, colour-blind-safe palettes, dark mode for charts, Stop in TOMO and Vision, a 'Based on' evidence line under every answer, and the sidebar grouped by job with Agentic BI renamed Deep Dive; the test suite grew from 90 to 322 blocks with ratchet lints on global assignment, silent handlers and file length
• Approved through the site SDLC as a non-GxP informational tool; a 16-artifact dossier — URS, architecture, test plan and scripts, traceability matrix, runbook, AI-governance statement — closed a platform audit finding