to adaptive learning
Automated metric tracking for the monthly operations scorecard, with an agent team that drafts the root-cause deep dive and signals the owner before a miss.
Biotech
Rare diseases and GMP manufacturing, where every batch matters.
External Ops
Runs a network of contract manufacturers across sites.
12 metrics
Quality, batch release and training, reviewed every month.
End to end
Designed and built it as the Insights & Analytics intern.
In a GMP operation a late deviation closure or a lapsed training record is a compliance risk, not just a red cell. Automation brings the numbers in; AI brings the why, early enough to act on it.
Every source lands in one governed layer and is calculated in one run. The main path ends in one drafted deck; a weekly branch signals owners, and their answers flow back into it.
deviations · CAPAs
batch release decisions · shipments
GMP training records · people
read-only from every source
Rate vs target, by site × month, with 13 months of history.
Copilot builds it from the metric store.
edits the read, then leadership signs off at the monthly review.
A Teams message or email: what's at risk, the context, and a drafted read.
confirms the cause, sets the action and an owner.
Read-only
Agents never write to a GxP system of record.
QA-confirmed definitions
Each metric's logic is signed off before it's automated.
Agents draft, people approve
Nothing reaches leadership without SME review.
Grounded and scored
Evals check every figure traces to the metric store.
The other 15% is on purpose: it's where people review, judge and sign off.
Hours went into pulling data and pasting it onto the deck.
The deep dive arrives drafted, and the team tests and refines it.
Misses surfaced at the monthly review, after the due date.
Early warnings reach the named owner days before the due date.
Days of coordination and manual pulls every cycle.
One scheduled source-to-deck run, refreshed more often.
Today a deviation is investigated, closed, then forgotten. If cause, owner, action and outcome are written back into one context layer, people and AI can ask:
The first step is already built. When an owner answers a weekly signal, their cause and action are captured. That is the start of the loop.
What it takes: an operating-model change, not another agent. Record the outcome where the work happens, and check effectiveness later.
Stack
Databricks (medallion, SQL) · Fivetran · Microsoft Fabric notebooks · Power BI semantic models (REST and DAX) · Delta tables · Power Automate · Copilot · agent evals
A note on what's shown
The flow follows the real build. Signal thresholds are illustrative; internal names and figures are left out. The 85% is the reported result.