Reconcile the development package against the GMP process, line by line
An agent that reads the development methods, specifications and process parameters against the ones actually running under GMP, and returns a line-by-line comparison: matched, differing, or present on one side only, with the source cited for every row.
The gap surfaces at engineering runs, six months after the handover.
The transfer package is handed over and it is thorough. Methods, specifications, the batch record master, the comparability protocol, development reports. Then six months later at the first engineering run the GMP method reads differently from the bench method it was built on, a limit was tightened by change control after the package was frozen, a raw material is specified at a different grade, and there is a parameter in the development work that has no equivalent anywhere in the batch record.
None of that was hidden. It was all in the documents and the spreadsheets. It was missed because reconciliation was two people reading two sets of material against each other and trying to hold four hundred parameters in their heads at the same time, under time pressure, at a point in the project where everyone wanted to move on.
The cost of finding it late is not the fix. It is the engineering run you burned, the schedule you lose, and the comparability argument you now have to make retrospectively.
Both packages, compared row by row.
The agent is pointed at the development package and the governed GMP systems. It extracts every method parameter, every specification line and every process setpoint from each side, matches them to each other across differing nomenclature, and returns a reconciliation table where each row is marked as matched, differing or single-source, with the origin cited on both sides down to the workbook and sheet or the system and record.
Matching is the hard part, and it is done conservatively.
The difficulty here is not reading the documents. It is deciding that development's "HMW post-capture" and the GMP specification's "HMW by SEC" are the same attribute, while not quietly matching two things that are genuinely different. The agent is built to be cautious in exactly that direction.
- The development package as it exists: process characterization workbooks, method qualification data, DOE runs, scale-up summaries and comparability work, read directly out of the working spreadsheets rather than requiring them to be migrated first.
- The governed systems on the other side: LIMS specifications and method parameters, MES process setpoints and in-process limits, and the controlled documents behind them.
- Attribute and parameter matching across differing nomenclature, with the basis for each match shown so a reviewer can overturn it, and the origin cited on both sides down to sheet or record.
- Unmatched items are surfaced, never dropped. Anything the agent cannot confidently pair is reported as unmatched for a human to adjudicate, which is the failure mode that matters: a missed difference is expensive, a flagged non-difference costs a minute.
- Change controls since freeze, checked against both sides, because the most common late surprise is a package that was accurate on the day it was written.
The agent finds and presents differences. It does not decide which differences are material to comparability, and it does not draft the justification for accepting one. A transfer lead adjudicates every row, and the unmatched items are deliberately put in front of them rather than resolved automatically.
From a reading exercise to an adjudication exercise.
- Two people reading two sets of material against each other
- Coverage depends on attention holding across several hundred parameters
- Change controls after package freeze are rarely re-checked at all
- Differences that are missed surface at engineering runs, at the worst cost
- Every parameter, specification line and method setting extracted from both sides and paired
- Unmatched items surfaced for adjudication rather than silently dropped
- Post-freeze change controls checked against both sides in the same pass
- The transfer lead spends their time deciding what matters instead of finding it
These figures come from my demo environment running on synthetic CMC data. They are not client results. The "today" column is what I watched teams actually do across HGS, GSK and Novavax; the "with an agent" column is measured on the demo, against a data model built to be representative rather than against your systems.
One side is spreadsheets and the other is validated systems, which cuts both ways.
Development content lives in working files and documents more than in systems. That is workable. The agent reads spreadsheets and native documents directly. What breaks it is a specification that exists only as a scanned signature page, a development file with three versions and no owner, or a controlled document system that cannot be queried programmatically. Where the receiving side is a CDMO rather than your own GMP floor, the same reconciliation applies and the binding question becomes what you are contractually able to access.
The Data Diagnostic establishes what is reachable on both sides before you plan around it, including the contractual question where a partner is involved, which is worth answering early.
The Data Diagnostic, including the price
Your IT and quality systems group will want the architecture, the read and write paths, the Part 11 position and the validation approach. That is all written down on the governance and validation page, in a form you can forward or print.
Send me a workflow.
Tell me the process that eats your team's week. I'll record an agent running it on your data model and send it back. No call required.
kyle@kylelangham.comNo form, no gate, no email capture.
Would rather talk it through? 30 minutes on your workflow, no pitch.