In meetings with Vitract.com (Canada and US) and PrecisionBiome.Eu (the EU) leadership this week, given the low cost of the new DeepSeek Large Language Model (LLM) model came up. Both of these firms are working on implementing their own variations of an expert system. At $6 million dollars, using DeepSeek open source model could easily done by venture capital back firms such as:
- Viome Life Sciences: – $175 million
- Viome’s AI system, called ‘Vie’ which uses machine-learning models for many chronic diseases. Evidence trails do not appear to be available.
- Seed Health: $44 milion
- Phylagen: $14 million
- Holobiome: $9 million
- BIOHM: $7.5 million
- Jona: $5 million
- Claims to use artificial intelligence to interpret microbiome data, analyze scientific literature, and provide actionable recommendations for probiotics, prebiotics, and dietary changes. Which AI models is not disclosed.
- BiomeSense: $3 million:
- HelloBiome: $4.8 million
- Claims to uses patent-pending AI-powered technology that uses supervised machine learning
- Enbiosis: less than $1 million
- Claims to have developed an AI algorithm that evaluates the relationship between gut bacteria and health parameters to create personalized nutrition strategies
The effectiveness is easily tested between the two, and made easier with a new addition on MicrobiomePrescription site. It allows you to take the high and low genus and ask these LLM AIs after getting the results of the Fuzzy Logic Expert System.

I have tried this with multiple samples and see that the LLM tended to rubber stamped answers similar to more intelligent influencers general advice ignoring the details. Below are some samples.
If you want to try it on your own favorite try the two below. The first one is a simple logic test: the same bacteria are listed as desired to both lower and to raise. All AI failed to see this issue!
What diet should I do to lower these bacteria: Acinetobacter, Anaerotignum, Barnesiella, Ruminococcus, Streptococcus, Subdoligranulum, Subdoligranulum and increase these bacteria: Acinetobacter, Anaerotignum, Barnesiella, Ruminococcus, Streptococcus, Subdoligranulum, Subdoligranulum?
The following has both different increases and decreases.
What diet should I do to lower these bacteria: Acinetobacter, Anaerotignum, Barnesiella, Ruminococcus, Streptococcus, Subdoligranulum, Subdoligranulum and increase these bacteria: Bombiscardovia, Faecalibacterium?
Results
Fuzzy Logic Expert System
ChatGPT
Perplexity
DeepSeek
This took many, many tries to get a response.

Black Box or Exposed Reasoning
Large Language Models hide their logic and are prone to Hallucination (artificial intelligence). Hallucinations makes them inherently unsafe for clinical use. On the other side, expert systems have their entire logic available. A good example is Example of Cross Validated Suggestions for Long COVID, where the full logic with links to studies is shown — with over 2100 links! An MD is easily able to evaluate suggestions and filter them by their own bias — or, be better informed on the current literature.
Additionally, counter indicated suggestions are included in the expert system evaluation. This is very unlikely with LLMs.
A more Complex Example
Rarely do we have such minor dysbiosis as shown above seen in most patients. Here is a more typical example.
What diet should I do to lower these bacteria: Acetobacterium, Acetobacterium, Acholeplasma, Acholeplasma, Alkaliphilus, Alkaliphilus, Alkaliphilus, Anaerotruncus, Anaerotruncus, Anaerotruncus, Anaerovibrio, Anaerovibrio, Butyrivibrio, Caldicellulosiruptor, Candidatus Amoebophilus, Coprobacillus, Dehalobacterium, Dolichospermum, Ethanoligenens, Fundidesulfovibrio, Fundidesulfovibrio, Hathewaya, Heliorestis, Heliorestis, Holdemania, Holdemania, Odoribacter, Odoribacter, Odoribacter, Odoribacter, Pseudoclostridium, Ruminiclostridium, Ruminiclostridium, Skermanella, Tindallia and increase these bacteria: Coprococcus, Dorea?
The above examples are both available on the Demo login of Microbiome Prescription.
For the last example, a full detailed report using the monte carlo model is attached below.
Bottom Line
The Fuzzy Logic Expert System used above have interesting statistics (here). Considering that it was produced by one person as a part time “hobby” over ~4 years should illustrate the feasibility of doing the expert approach. Mind you this person had the right skills:
- Taught AI at University for a few years
- Worked professionally in AI for firms such as Microsoft, Verizon, and Amazon
- World class programming skills (including white papers for Microsoft and others)
- Taught science at High School and Colleges (and has read medical papers since 15 y.o.)
- Is high functioning Autistic — allowing prolonged focused concentration on issues
This post indicate future trends:

IMHO, it is morally and professionally irresponsible for suggestions / therapies to be made without all of the evidence that the suggestions and/or therapies is based on to be available in a human (MD) readable format. To the best of my knowledge, none of the vague AI claims above provide that to their customers. “Machine Learning” is a black box. Claiming AI is often a marketing strategy that border on fraud.
Recent Comments