Benjamin Rizkin, PhD in chemical engineering. Inventor on six patent filings across ten jurisdictions, nine publications, first author in Nature Machine Intelligence. Fluent in reactors, regulators, and the space between them.
From a route on a whiteboard to a process a regulator will license. The practice of Benjamin Rizkin, PhD, chemical engineer, builder of complicated things that work.
Benjamin Rizkin, PhD in chemical engineering. Inventor on six patent filings across ten jurisdictions, nine publications, first author in Nature Machine Intelligence. Fluent in reactors, regulators, and the space between them.
CMC and technical operations across the whole life of a product: route, preclinical, IND, Phase 1 through 3, validation, filing, inspection, launch. Small molecules through living cells, and the analytical, supply and quality systems around them.
Acoustics and signal processing, computational fluid dynamics for , photography above and below the waterline, vintage Swedish automobiles, jazz.
Most CMC programs run on a handoff. One person knows what the process will actually do. A different person owns the schedule, the budget and the filing. Quarters disappear in the gap between them, because the plan was never built on the science and the science was never told what the plan could absorb. I have worked both sides of that line on the same products. Everything below is work I have done, not services I am offering to learn.
The impurity that only appears at scale, and the fate and purge model that should have predicted it. A method that drifts, and the analytical lifecycle procedure that catches the drift before a lot fails rather than after. Nitrosamine and elemental impurity risk assessments, and knowing which of the two is actually your problem. Comparability after a process change made during pivotal trials, which is the version everyone is afraid of. Chain of identity and chain of custody, without which an autologous lot is not releasable at any level of potency. Assays that have to be agreed across analytical, QC, quality and regulatory before they mean anything. A single-site API, and the second source qualified two years before anyone asked for it. Cold chain and shipping validation for a product whose shelf life after thaw is measured in hours.
There is no CMC function yet.
A candidate, a target product profile, and nobody whose job this is. I have built the whole function twice: strategy, the external network, the quality system, the budget and the filing content. One emerging biotech from nothing to two programs at IND readiness and six molecules through preclinical. Another from a company bench to commercial ready.
The lots are not releasing.
The process runs, mostly, and the deviations are winning. Nothing here is a chemistry problem by the time it reaches you. A Phase 3 autologous site taken to eight times the lot output it started at, with remanufacture nearly eliminated and process deviations halved.
The filing and the inspection are coming.
Module 3 that holds, a control strategy that survives a question, and readiness built as an evidence system rather than a room of filing cabinets nobody has opened since the last one. Module 3 sections in NDA filings. CMC accepted cleanly by FDA, EMA and MHRA. A pre-license inspection program under RMAT.
Nobody can say what the critical path really is.
Every function has a plan and none of them are the same plan. What is missing is not software, it is a governance structure where the decision lands where the accountability already sits. A four-tier governance design directing portfolio, technical risk and resourcing across eighty workstreams and sixteen departments.
It is not a drug, and the regulator is just as hard.
Combination products, devices, and the industries next door. The authority changes and the discipline does not: a basis agreed up front, evidence built to match it, and a record that survives somebody reading it years later. A drug-device combination product. An aircraft company, where the same argument is made to a different agency.
Nobody here can read a quality agreement.
The quiet fear of every company without a plant: that a vendor is taking advantage and there is no one inside who can tell. I have sat on both sides of that contract, and been the person standing in the plant. Selection, contracts, quality agreements, vendor qualification, governance, second sourcing, person-in-plant.
Route, scale-up and formulation, with a control strategy argued from mechanism rather than asserted.
Methods, the procedure lifecycle, and specifications set with QC and quality rather than handed to them.
ICH Q3A through Q3D and M7, fate and purge, nitrosamines, and which of them is genuinely your problem.
Chain of identity for one-patient lots, conjugation ratio, sterility assurance, device constraints.
The integrated plan and critical path, with decisions landing where the accountability sits.
Selection, contracts and oversight, then second sourcing, cold chain and the contingency behind both.
Deviations, root cause and CAPA, and a straight answer about supply risk before it becomes an interruption.
PPQ, commercial transfer, Module 3, and a cost of goods decided at route years before anyone reports it.
Programs. Adrabetadex, a sterile intrathecal therapy for Niemann-Pick Type C, owned end to end through three CDMOs. Rilparencel, an autologous cell therapy in Phase 3 under RMAT, through pivotal execution into BLA and pre-license inspection readiness. Venclexta, Rinvoq and Vyalev on the approved side. Navitoclax and ABBV-154 through late-stage validation and PPQ. ABBV-155, ABBV-484 and ten further molecules from preclinical into the clinic.
Breadth. Small molecules, peptides, antibody drug conjugates and bioconjugates, biologics, autologous cell therapy, and a drug-device combination product. Sterile parenteral and intrathecal, oral solid dose including granulation and blending, topical, spray drying. Beyond the dosage form: analytical development and the method lifecycle, international supply chain and distribution, quality systems, a commercial facility from design through qualification, and the digital backbone a modern site releases through, ERP, LIMS, MES, and chain of identity and custody. What transfers between an intrathecal cyclodextrin and a living cell therapy is not the chemistry. It is the system around it.
Also. A dual-facility commercial build guided from design through qualification, sized for roughly twelve times the clinical run rate. An AI and machine learning platform for molecular design and process optimization, selected for the NVIDIA Inception accelerator. Six patents and applications across ten jurisdictions, covering enzymatic cyclodextrin production and modular manufacturing. Nine publications, first author in Nature Machine Intelligence.
Engagements, through Rizkin Labs LLC.
Fractional head of CMC, an external network built or repaired, filing and inspection readiness, technical diligence for a raise, or a site that is not running right. Scoped to the problem, with a date on it and success criteria we agree at the start and then meet.
Or the permanent seat.
Senior technical operations and CMC leadership. Winston-Salem, and willing to move. If you are hiring rather than retaining, the same evidence applies and the conversation is shorter.
A CMC function that does not exist yet. A process that works in development and not in production. An external network with nobody watching it. A filing or an inspection with a date on it. A cost of goods number that has to come down, whether the process is young enough to redesign or already running in Phase 3.
Viral vector gene therapy, mRNA and vaccines are not in my record. I could read my way in over a long weekend, and you would be paying for the weekend. Ask me to run your clinical strategy and you will get a very confident answer from a man who should not be giving one. If that is the problem, I would rather say so on the first call than three invoices in.
Nearly every company in this industry now has a data lake and an AI initiative. Far fewer have a single model anyone will let near a batch disposition, a specification or a submission. The gap is not modeling talent, which is cheap and getting cheaper. It is that a model used for a decision a regulator can ask about has to be justified the way an analytical method is justified, and the people who build models have usually never had to defend one. I have trained machine learning models for molecular design and process optimization, and I have sat across a table from an inspector explaining a computerized system.
A model that predicts which lots will miss release, which fails the first time because the records it learned from do not agree with each other about what a lot is. An attribute you cannot measure while the process is running, inferred from three you can, and the evidence package that lets you act on the inference rather than merely present it. Ten years of deviation text, clustered, which usually shows the same three root causes rewritten four hundred times. Release testing that takes a week because the panel was designed sequentially and nobody has ever modeled the queue. A design space fitted rather than gridded, so the study that wanted thirty runs takes twelve. Chain of identity and chain of custody, which stop being a paperwork problem and become a data problem the moment there is more than one patient. A forecast that states its interval instead of drawing one confident line into next year.
What a site actually produces, and where it dies: historians, LIMS, MES, ERP, batch records, instrument files nobody can join. Data integrity first, because a number you cannot trace back to the instrument that made it is not evidence.
Batch genealogy, the real drivers of yield and cycle time, deviation clustering, cost of goods by step. The test is whether somebody does something differently on Monday.
Property and route prediction, design space and study surrogates, soft sensors for what the line cannot measure, forecasting for supply. Fitted to your data, not to a benchmark.
Deviations, CAPAs, protocols, SOPs, health authority correspondence. Retrieval that quotes the paragraph it came from and links to it, so the reviewer checks the source instead of trusting a summary.
GAMP 5, Part 11 and Annex 11, and the credibility argument a regulator now expects for a model in a stated context of use. What an intended use costs you in evidence, settled before you build.
The thing that has been a spreadsheet with four owners for six years. Small, fast, hand-written software that does one job and can be read by the person who inherits it.
First author in Nature Machine Intelligence, nine publications, including work on word embeddings. An AI and machine learning platform for molecular design and process optimization, selected for the NVIDIA Inception program. The digital backbone of a commercial site, ERP, LIMS, MES and chain of identity and custody, specified and stood up rather than bought and hoped for. And the Bench on this site: fourteen working instruments, hand written, no frameworks, which is the smallest honest demonstration of custom I can hand you without an invoice.
And the rules. The EU's draft Annex 22 on artificial intelligence, the rewritten Annex 11, and the FDA's credibility framework are all still drafts, and between them they decide what a model is allowed to touch. I wrote up where each one stands and what to build before any of them is final: AI under GMP.
A data estate nobody can join. A model that works in a notebook and has no route into a GxP system. A decision being made on a spreadsheet that deserves a model, or a model about to be built that deserves a spreadsheet. An AI claim in a diligence deck that somebody needs to check. A tool that should exist and does not.
Training a foundation model of your own. Replacing a QC lab with a prediction. An AI strategy deck with no working system at the end of it. If what you need is a number on a slide to raise against, I am the wrong person, and you will hear that on the first call rather than in month three.
Long-form field guides on pharmaceutical CMC, Quality by Design, analytical lifecycle, and the program leadership that carries complex therapies to patients. Written to be used rather than cited, each with reference tools that run at your desk.
The referenceThis industry runs on abbreviations and four-character document codes, and almost nobody holds all of them at once. Every ICH guideline, the parts of 21 CFR that actually come up, the EU GMP annexes, the USP chapters, the PDA technical reports and nearly seven hundred acronyms, one thousand and six entries in one searchable box. Type any of it, or any part of what it means, and it resolves.
Open the pharma acronym finder →I co-founded SkyNext Aviation and exited. The configuration was a box wing: Prandtl's best wing system, the closed lifting arrangement he proved in 1924 has the least induced drag of anything you can build at a given span and lift. Aldo Frediani's group at Pisa carried it from theorem to aircraft, and the EU's PARSIFAL project took it through a full transport-category design. We were building it small first, and building it by printing it. An aircraft and a medicine are both things you may not sell until an authority agrees you have shown your work. Different regulator, same discipline.
Keys, with the panel or the model under the pointer: W/S throttle · R/F tunnel speed · ←/→ roll · ↑/↓ pitch · A/D rudder. The model holds whatever attitude you set, so you can study one; center resets it. Drag the model to look around it.
The flow is solved, not drawn: both wings cut into panels, each carrying a horseshoe vortex, velocity by Biot-Savart, streamlines integrated through the result. The airframe is the CAD mesh, decimated, lit and depth sorted so it occludes both the flow behind it and its own far side. The quadrant flies it: the model turns inside a tunnel that stays where it is, and the field is re-solved a slice at a time so the picture never blinks. The four wing rotors and the ducted fan are actuator discs, which is why the wash leaves in a corkscrew and runs brighter than the air around it. All of it public science: Prandtl in 1924, Pisa since the eighties, PARSIFAL in the open under Horizon 2020. Anyone can draw a box wing. What a closed wing lets you do to the structure, once you are no longer obliged to build it the way wings have always been built, is not on this page. Ask me on a call.
SkyNext was never one person. Engineers, advisors and friends put real work into it, for no certain return, and the parts worth being proud of have their fingerprints on them. My thanks to all of them.
Engagements through Rizkin Labs LLC, or the permanent seat. Either way the first conversation is free and I will tell you if I am the wrong person for it.