From weeks to minutes: Automating Experiment Design with AI for Bioprocessing Scale-Up

HIGHLIGHTS

• From weeks to minutes: experiment design accelerated by as much as 90%
• A complete prompt-to-virtual-run recipe design cycle
• An AI-enabled DoE copilot for bioprocess recipe design and acceleration
• Self-learning system improving with every run
• LLM-powered AI Assistant built into the customer platform
• Delivered under strict regulatory and IP constraints

OVERVIEW

The Client is a cloud-connected bioreactor and bioprocessing software manufacturer that helps life sciences companies optimize their manufacturing processes to grow living cells and biomolecules faster and more efficiently.

At the core of their offerings is the Cloud Bioreactor Lab — a high-throughput facility where the client runs experiments on behalf of their customers. Data is generated in the laboratory, streaming results back in real time so customers can monitor and analyze runs from anywhere.

By providing both the physical infrastructure and the operational expertise, The client expands its customers' lab capacity, enabling them to run more experiments than would otherwise be possible.

THE CHALLENGE

As the Client scaled its customer base, it needed a matching software solution to accelerate the experiment design process while maintaining a high level of accuracy. The core challenges that the client wanted to resolve included:

• Time-consuming manual experiment design. Every bioreactor run begins with creating a recipe, the defined physical parameters (temperature, pressure, moisture, etc.) that govern the process. Previously, there were no ready-to-use libraries or AI tools for this process; every recipe was designed manually by researchers working with client's clinical chemists.
• A promising start that stalled. The Client had already attempted to build a generative AI assistant using their collection of historical recipes and experimental data to take a natural-language input — such as growing a specific molecule in a defined timeframe — and automatically generate a deployable recipe. The initiative showed real promise but stalled without dedicated software engineers to carry it forward.
• Tight IP and data security constraints. Strict industry compliance and regulatory requirements and proprietary customer experimental data placed firm limits on how historical runs and recipes could be used and shared for AI training — ruling out an unrestricted, at-scale approach and requiring a carefully governed, tightly scoped way of putting that data to work.

The Client approached Kanda Software as an AI development partner with in-depth domain expertise in Life Sciences and Biotech to make experiment design faster and more predictable — reducing errors, cutting costs, and building a system ready to scale.

SOLUTION

Client's scientists already had a tool that let them describe an experimental goal in natural language and receive a first-pass design back. Kanda's job was to extend the same natural-language interaction across the rest of the process, turning scientists' prompts into a fully structured, statistically sound experiment design.

To do that, Kanda trained the underlying model on a representative set of client's own historical recipes and experiment plans as well as public scientific repositories, in close collaboration with client's team, so the outputs reflected real bioprocessing practice rather than generic assumptions.

In parallel, the Kanda team established statistically grounded parameter ranges for recipe recommendations. The team ran a set of experiments that enabled them to finally complete the full cycle — from natural-language prompting through parameter selection and recipe generation to a virtual run in the Cloud Bioreactor. The results enabled Kanda to further refine and improve the process of recipe generation.

RESULT

A scientist describes what they're trying to achieve, and the system proposes a design defined in proper statistical terms — which parameters to vary, over what ranges, and how they interact with each other — instead of a rough outline the scientist still has to formalize by hand.

From there, the same information keeps moving forward automatically. The variables that define the experimental design flow directly into recipe generation, and the recipe's parameters flow directly into experiment-plan generation, so nothing has to be manually re-entered at any step. At each stage, the scientist reviews what the AI proposes and can ask for changes before moving on, always keeping a human decision-maker in the loop.

Every recipe is validated against the client's platform before it's saved — so what scientists get back isn't just a plausible-looking suggestion; it's something confirmed to work within their actual system.

Alongside the DoE Co-Pilot, Kanda also implemented "Talk to Your Recipe" — a GenAI interface that lets operators configure existing device parameters by natural language prompts, without searching through documentation.

The result is a production-ready AI Assistant, built into Client's software platform, that takes natural-language inputs and fully automates experiment-plan generation.

HOW THE AI IS KEPT TRUSTWORTHY

The system's trustworthiness rests on four enforced principles: grounding, verifiability, constraint, and human oversight. Every output is traceable, validated, and subject to expert review before use:

• Grounded in client's data — retrieval runs over company's historical experiments and methods, so every recommendation is anchored in real evidence rather than generic assumptions
• Agentic workflows — the system drafts a design, checks its own work against statistical rules, and revises — the way an expert scientist would iterate
• Deterministic guardrails — hard science and validation logic wrap the model, so every output is constrained and independently checkable, not just a plausible-sounding answer
• Human-in-the-loop — the AI proposes; a scientist accepts, edits, or rejects every step, so people stay firmly in control of the decision

BUILT FOR REGULATED, IP-SENSITIVE ENVIRONMENTS

In Life Sciences, experimental data carries a dual sensitivity: it is both commercially proprietary and, in regulated workflows, subject to audit. This engagement was architected accordingly, with data governance treated as a design constraint from day one rather than a compliance afterthought:

• Data segregation approach - customer data isolated per environment. No cross-tenant training
• GxP, FDA 21 CFR Part 11, and HIPAA requirements are built into Kanda's delivery process by default.
• A human review remains in the loop at every stage of a recipe generation process, with every output validated against the client's own platform before use.

OUTCOMES

• ~90% faster experiment design.
• Rising success rate of new runs. The system becomes more accurate with each cycle, increasing the success rate of new runs as it learns from more data.
• Delivered under strict regulatory and IP constraints, without relying on unrestricted use of proprietary customer data.
• Foundation for full-scale experiment orchestration. With the complete recipe library accessible through one AI interface, the client can coordinate simultaneous bioreactor processes and scale as its customer base grows.

TECH STACK

The solution is powered by enterprise-grade large language models (Gemini and Claude), with a React front end and Python back end deployed on AWS — giving the client a secure, scalable foundation that grows with its customer base without introducing new data-handling risk.

Ready to Accelerate Your Own Bioprocess Cycle?

If your team is losing weeks to manual experiment design or sitting on proprietary data you can't risk exposing to an off-the-shelf AI tool — Kanda's engineering team can help you scope a similarly governed approach.

Reach out to our AI engineering team: contact@kandasoft.com