I am A.B. Modi. I build software that shows its working.
Ligant is where I build and deploy open source tools and systems for cell and gene therapy research and development. They are deterministic, they run in your browser or on your own hardware, nothing you type into them leaves your machine, and they are built to refuse rather than return a confident number the inputs do not support.
A.B. Modi LLC is where I help companies validate the AI systems they already have and build the ones they cannot buy: proprietary models and deterministic AI systems trained on their own data and running on their own infrastructure. I work alongside pharma and biotech IT and quality teams, and a good deal of that work is coaching rather than delivery.
The two are the same problem seen from two ends. A result you cannot inspect is a result you cannot defend.
Ligant
I founded Ligant to build and deploy open source tools and systems that advance cell and gene therapy research and development. We started with the parts of that work where a number decides something, and where the number can be wrong in ways that look right.
Four tools are live at benchtools.ligant.ai: an Antigen Density Calculator, a Molarity Converter for biologics, an Antibody Titration Planner and a Dilution Planner. Each does one job. Each states the assumptions it applied, echoes the inputs it used, and names the conditions under which it will not return a value at all. They run entirely in the browser, need no account, and are published under the Apache License 2.0 with the source public.
The position is narrow on purpose. Everyone already has a calculator. Almost nobody has a checker. The arithmetic in these tools is arithmetic a scientist can read. What they add is the set of checks that catch a calibration or a declaration that has quietly made the result meaningless, and the refusal to print a confident figure when those checks fail.
For regulated environments
For organisations that cannot use a public site, Ligant deploys the same tooling privately on the customer’s own infrastructure: reference databases, connected workflows where a flag raised in one tool carries into the next, local language models, and the validation package a GxP workflow requires. Ligant produces the evidence and the package. The customer’s quality unit executes and approves it.
The free tools are not a trial of that, and the private deployment is not a better edition of the free tools. They are different things.
This is Ligant tooling, deployed for you. Validating a system you already own, or building one to your own specification, is the other practice below.
A.B. Modi LLC
Separate practice, separate problem. Ligant publishes into the open. This work runs the other way: into organisations that cannot send their data anywhere and still have to defend what their AI does.
Validating AI systems. I help quality and IT teams write an AI validation plan and then hold a system to it, under Annex 11, cGMP, and the draft Annex 22 on artificial intelligence. Annex 22 is not settled law yet, which is the argument for starting now rather than after it lands. Its draft scope points somewhere useful too: static models with deterministic outputs sit inside it, while continuously learning models and large language models are kept out of critical use.
Building proprietary AI. Custom language models and deterministic AI systems, trained on a client’s own data and running on their own infrastructure. Two things usually drive it. The data cannot leave, and a system whose behaviour you can bound is a system you can validate.
Responsible AI. The work is aligned to the ten joint FDA and EMA guiding principles on the use of AI in drug development. Those principles are high level rather than binding, so I say aligned to and not compliant with. The second phrase means something specific and it does not apply here.
Coaching. A fair amount of this is not delivery at all. It is sitting with IT and QA while they work out what validation means for a system that will not do the same thing twice, and where the honest answer is usually to narrow what the model is permitted to decide.
How I work
Four positions carry across both halves of the work.
Show the working, including the parts that failed review. The Antibody Titration Planner was built against a formal specification that our Chief Scientific Officer rejected three times. The findings included a flag that stated a concentration ratio backwards and a layout that let a scientist read twelve numbers with none of the conditions visible. The specification and those findings are public alongside the tool. That is more persuasive than anything I could assert about quality.
Refuse before you guess. A wrong number that looks right costs more than no number. Where a value cannot be computed from what was actually declared, the tools say so and say why, instead of printing a plausible figure with a caveat underneath. Each tool also lists what it cannot detect. The titration planner names fourteen such failures on its own page.
Keep the AI out of the arithmetic. The bench tools contain no model and perform no inference. AI was used to build them. It does not run inside them. I am specific about this because it is the reason a result can be cited, and because the distinction matters more, not less, in the systems where there is a model.
The scientist decides. These tools compute a conversion or a plan. They do not read the data, pick the point, or stand in for the person who knows the biology. That is a design position rather than modesty.
Get in touch
The single most useful thing you can send me is the calculation you keep redoing by hand, and the reason the spreadsheet version of it worries you. That is how the next tool gets chosen.
- Consulting, private deployments, validation work
- ab@abmodi.ai
- Ligant, anything else
- hello@ligant.ai
- The tools
- benchtools.ligant.ai
- Source
- github.com/abmodi-ai
- linkedin.com/in/abmodi-ai
The tools are free and will stay free. There is no account and no sign up. Nothing you type into them leaves your browser; page visits are counted through the hosting provider, because without a login there is no other way to know whether any of this is useful.