CMC Decision Platform
A concept design that puts CMC, clinical, and regulatory inputs into one decision memo with an audit trail.
Interactive walkthroughThis is a scripted walkthrough with synthetic data. It shows how the tool works without real program data.
Phase gate
Select a gate to see the memo that the platform writes for it.
Decision memo
- RecommendationGo with conditions
The filing can keep its May 2027 date if the remaining stability data and the second-site report arrive before the dossier lock.
Combined inputs
- CMC readiness. Validation batches 3 of 3 passed. The team has 12 of 18 months of stability data.Open item
- Clinical. The study met its primary endpoint. The final study report is signed.Ready
- Regulatory. The team sent the Pre-BLA questions. The team expects the agency to ask for 18-month stability data at filing.Open item
- Supply. The main site is qualified. The comparability report for the second fill site is pending.Open item
Conditions
- The team adds the 18-month stability results before the dossier lock in March 2027.
- Quality approves the comparability report for the second fill site.
- If one of these items is late, the memo goes back to the forum.
Why I built it
At each phase gate, a CMC decision needed inputs from many functions at the same time. These included clinical dosing assumptions, regulatory precedents, commercial supply limits, and process and analytical data. In my experience, no single system connected them. Teams made decisions without full visibility across functions and without a traceable record of the reasoning.
What it does
- Puts the CMC, clinical, regulatory, and supply inputs for one phase-gate decision into one view.
- Drafts a decision memo that lists the options, the evidence behind each, and what is still unknown.
- A named subject-matter expert reviews and signs off. The audit trail keeps every edit.
- This tool is a concept design with synthetic data. It is not in use on a program.
- What goes in
- CPP and CQA data, clinical PK/PD and dosing signals, regulatory guidelines and precedents, commercial supply and CoGS assumptions, structured analytical data, and past program precedents.
- What comes out
- A decision memo with the options and the evidence for each, a risk view for the gate, and the evidence gaps to close before filing.
- What it changes
- A phase-gate meeting starts from one shared view of the evidence and the open questions. Each function does not bring its own slides.
How it stays inside the rules
- What it is for
- Phase-gate decision support across functions. The tool gives a governance forum a reasoned recommendation with a traceable chain of arguments. Of all the tools on this site, this tool has the highest consequence, and it has the strictest controls.
- What it never does
- Explicitly out of scope for critical GMP use. This tool is not permitted near batch disposition, specification decisions, or process control. Its design does not include these uses. It gives information for strategy decisions that a governance body makes.
- Who decides
- The output is a reasoned recommendation to a phase-gate forum, and each input is traceable. The forum makes the decision and records it in the minutes. The model never makes the decision.
- Risk level
- High under FDA's draft AI framework (how much the output influences a decision × how serious that decision is)
- Guidance it follows
- FDA draft guidance · EMA reflection paper · ICH Q8 to Q14
What already exists
- No direct incumbent: Each function has its own point solutions (RIM, LIMS, PLM, clinical, commercial forecasts), but no reasoning layer connects them
- The gap it targets
- Phase-gate CMC decisions depend on four inputs at the same time: clinical dosing assumptions, regulatory precedent, supply economics, and process data. The reasoning that connects these inputs is only in a slide and in the memory of one person. This tool is a concept architecture for that layer. It has a High model risk rating because the consequence is strategic.
- What it is not
- It is not a shipped product, and this site does not show it as one. It is an architecture with a stated credibility case. Today, that is the honest state of the art for AI that supports decisions across functions in this industry.
Do you build tools like this?
I am glad to compare notes: what worked, what a reviewer did not accept, and where AI should stay out of the decision.