From microbial discovery to automated production

I want to manufacture a living microbial crop-protection inoculant and build an automated workflow that measures whether each batch remains viable and biologically active.

1. What I want to make—and why

I am building Elemin around two connected systems: MAIA (Microbial AI for Agriculture), an AI system for discovering beneficial microorganisms, and Athanor, an automated manufacturing system for producing them.

MAIA generates candidates from genomic evidence. Turning a promising candidate into a useful product introduces another set of questions: Can it be grown reproducibly? Does it retain its useful function after production? How long does that function last? Which measurements should an automated production system trust?

For this course, I want to make small research batches of a single-strain liquid bacterial inoculant intended for biological crop protection. Each batch would have a documented strain identity, viable concentration, functional measurement, and production history.

The long-term audience is growers, including producers of tree crops such as coffee and cacao. The immediate output is a characterized research prototype and a production workflow that can later be applied to MAIA-selected candidates.

My central question is:

Can we manufacture a beneficial microorganism reproducibly while preserving the biological function that made it worth selecting?

This connects what I am already building in software with what I need to learn at the bench.

2. What living system I would use

My proposed starting organism is Bacillus velezensis FZB42, a well-characterized research reference. Its documented products include bacillomycin D, fengycin, and surfactin, providing a biological basis for investigating crop-protection activity. Koumoutsi et al., 2004

DSMZ lists FZB42 as Risk Group 1. It also explicitly restricts commercial use and transfer to third parties, so I would use it as a research reference under the applicable terms. A subsequent Elemin product would require a strain with suitable commercial rights. Final organism selection and sourcing would be agreed with the host lab. DSMZ strain record

The high-level process is:

Identified starter culture + nutrients + water + oxygen → controlled microbial growth → recovery and liquid formulation → measured living product.

The microorganisms are themselves the manufactured product. I would measure viable output and retained biological activity separately, because a dense culture does not necessarily make an effective inoculant.

Once the reference workflow works, a useful comparison would be Pseudomonas protegens Pf-5, a documented biological-control research strain with a different production and formulation challenge. It would serve as a reference comparator; I would not assume which organism performs better. ATCC strain record

The subsequent discovery step would be to apply the established workflow to an accessible MAIA-selected strain whose exact genome and culture identity have been matched.

3. How I would use automation

I would build the first Athanor prototype around programmable liquid handling, recorded culture measurements, and software that connects every result to its strain and production recipe.

An Opentrons OT-2, subject to node availability, would automate repetitive preparation, transfers, dilutions, and assay setup. These are established capabilities of the platform. Culture incubation and measurements would use the host lab’s available equipment, with manual transfers recorded where instrument integration is unavailable. Opentrons OT-2

The workflow would have four parts:

  1. Prepare and identify. Assign strain, recipe, batch, and sample identifiers; automate suitable preparation steps.
  2. Produce and record. Collect available growth and process measurements alongside instrument actions and manual interventions.
  3. Measure quality. Determine viable concentration and activity in one quantitative, instructor-approved antagonism assay, with appropriate controls.
  4. Choose the next experiment. Compare batches by usable viable output, retained activity, reproducibility, and hands-on effort; use the results to select the next production condition or candidate.

My main engineering deliverable would be an executable workflow and a batch report linking what was intended, what happened, and what was produced.

If time and equipment permit, I would investigate whether inexpensive optical measurements predict independently measured batch quality. That relationship would need calibration and testing; fluorescence or growth alone would not be treated as proof of potency or absence of contamination.

4. How much I would make—and how long it would take

Course prototype

My initial target is three independent batches at approximately 100 mL working volume each. This is enough to begin measuring production variability while keeping the work manageable.

For planning, I assume:

  • Harvest concentration: 10⁹ colony-forming units per mL (CFU/mL).
  • Recovery into the research formulation: 50%.

Estimated recovered output per batch:

100 mL × 10⁹ CFU/mL × 0.5 = 5 × 10¹⁰ CFU

Three batches would therefore yield approximately 1.5 × 10¹¹ CFU, before allocating material to measurements and storage studies.

These are planning assumptions, not demonstrated performance for FZB42. As an order-of-magnitude reference, a published manufacturing study of a different strain, B. subtilis CW-S, reported approximately 10⁹ CFU/mL at laboratory scale and 2 × 10⁹ CFU/mL at industrial scale. Transfer to my strain and equipment must be measured. CW-S manufacturing study

Assuming a 72-hour production cycle, including preparation, growth, recovery, and turnaround, three sequential batches would require approximately nine days of production, followed by completion of quality measurements. I would allocate four to six weeks for setup, initial qualification, repeat batches, and short-term stability observations.

Meaningful manufacturing scale

For an industrial thought experiment, I would consider one application across 1,000 hectares.

I assume an illustrative requirement of 10¹³ CFU/hectare. This is a capacity-planning assumption; an effective agricultural dose would require separate plant and field evidence.

Total required viable output:

1,000 ha × 10¹³ CFU/ha = 10¹⁶ CFU

At the assumed harvest concentration and recovery:

10¹⁶ CFU ÷ (10⁹ CFU/mL × 0.5) = 20,000,000 mL = 20,000 L of culture broth

With a 1,000 L working-volume reactor, this requires 20 batches. At an assumed 72-hour cycle, production would take:

  • One reactor: 60 days.
  • Four parallel reactors: 15 days.

Final quality testing and distribution would add time. Lower recovery or loss of viability before application would increase the required production.

At this scale, automation would shift from laboratory liquid handling to instrumented bioreactors, controlled material transfers, process monitoring, and batch-quality decisions.

The calculation shows why I need to measure both yield and retained function. Athanor’s proposed production close to use also makes the delivery interval important: the best organism for fresh local supply may differ from the best organism for prolonged storage.

5. The three biggest unknowns

Does useful activity survive production?

I need to establish whether batches retain measurable biological activity after growth, recovery, and formulation. An antagonism assay would establish activity under its test conditions; demonstrating crop protection would require subsequent plant studies.

What viable yield and stability can I achieve reproducibly?

I need to measure variation between independent batches and changes during storage. These results will determine how much usable product reaches the intended point of application.

Which measurements can support automated decisions?

I need to learn whether accessible process measurements predict independently measured quality, and whether the workflow remains useful when moving from a reference strain to a MAIA-selected candidate.

What success would look like

A successful course project would deliver characterized living research batches, a repeatable automated workflow, and a dataset connecting strain identity, production conditions, viability, and biological activity.

Those results would give MAIA experimental evidence about its candidates and give Athanor a first tested production workflow. Lab access would allow me to connect genomic predictions to measurements that software alone cannot provide.

AI assistance

I used Claude via Claude Code and OpenAI Codex for research, drafting, and critical review. Representative prompts included:

  • “Now, based on what I am doing with MAIA, I want to propose growing the most beneficial to the mission of MAIA microbial products. Analyze everything and help me narrow what I should do.”
  • “While the point of taking this course for me is learning automation and biomanufacturing, I want to biomanufacture useful things for Elemin, what should the homework #1 proposal be?”
  • “Review Claude session ‘66. HTGAA microbial products strategy’ and let me know if I should consider anything else.”

Published findings are linked above; proposed production quantities, recovery, application requirement, and scheduling are identified as assumptions.