GutIQ
GutIQ Research · Medical AI

Meet CEREBRA

A mixture-of-experts medical model. Trained only on medicine.

Cerebra is the day-to-day clinical assistant inside the GutIQ stack — eleven specialist sub-models routed through a single conversation, every answer grounded in primary literature, and a strict guardrail that turns away anything outside medicine. Built for patients on a programme; useful for anyone asking a careful question.

11
specialist experts
9.2M
curated abstracts
15yrs
internal clinical data
0
PII in training
01 · The Idea

One general model can't think like eleven specialists.

General-purpose chatbots are trained on the open internet — a corpus heavily weighted toward forum opinion, marketing copy, and old wives' remedies. They will happily compose a confident paragraph about a gut symptom while citing nothing more rigorous than a wellness blog. That is not what medicine looks like.

Cerebra is the opposite. It is a mixture-of-experts (MoE) system — a small constellation of specialist sub-models, each fine-tuned on its own narrow slice of the literature, with a router that picks the right experts for the question in front of you. You see one continuous conversation; underneath, the question on slow transit goes to the motility expert, the question on bloating after antibiotics goes to the dysbiosis and pharmacology experts. The result is an assistant that reasons like a tumour board, not a search bar.

02 · Architecture

Eleven experts. One conversation.

Each expert is fine-tuned on its own domain corpus. The router model — itself fine-tuned on 60,000 labelled clinical questions — chooses which subset to activate. Most queries route to two or three.

e_01
Functional GI

IBS, dyspepsia, slow transit, pelvic-floor dysfunction. ROME-IV grounded.

e_02
Microbiome

Dysbiosis, SIBO, post-antibiotic recovery, fermentation patterns.

e_03
Nutrition Science

FODMAP biochemistry, fibre fractions, micronutrient interactions.

e_04
Pharmacology

Drug-induced GI effects, PPI tapering, prokinetics, antibiotic stewardship.

e_05
Gut–Brain Axis

Vagal tone, stress–symptom coupling, visceral hypersensitivity.

e_06
Diagnostics

Stool, breath, blood marker interpretation; when to escalate to imaging.

e_07
Symptom Trajectory

Reads your daily log series. Detects rebounds, plateaus, false-positives.

e_08
Sleep & Circadian

Migrating motor complex, shift-work patterns, melatonin–gut crosstalk.

e_09
Exercise Physiology

Intensity–GI symptom curves, hydration, race-day protocols.

e_10
Red-Flag Triage

Always-on. Flags blood, weight loss, nocturnal pain, fever before answering.

e_11
Plain-Language

Translates expert outputs into a single voice. Calibrates reading level to you.

router
The Conductor

Reads your question, picks experts, weights their answers, returns one response.

03 · Training Corpus

What we feed it. What we don't.

Cerebra is fine-tuned exclusively on medical literature and our own de-identified clinical record. No Reddit. No Quora. No wellness influencers. No raw web scrape.

PubMed Central
6.4M open-access full-text articles

GI, microbiome, nutrition, and adjacent specialties. Filtered to peer-reviewed primary research, systematic reviews, and meta-analyses.

open data
Cochrane Library
All gastroenterology and nutrition reviews

The gold standard for evidence synthesis. Weighted heavily in expert e_06 (Diagnostics) and e_04 (Pharmacology).

licensed
HuggingFace
MedQA · MedMCQA · PubMedQA · MMLU-Med

Open-source medical Q&A and reasoning datasets used for instruction-tuning. Lets Cerebra answer in the way a clinician would, not the way an essay would.

open data
ICD-11 · SNOMED CT
Clinical ontologies

Standardised terminology so symptom descriptions map cleanly to recognised conditions, not improvised labels.

licensed
NICE · BSG · AGA
Clinical practice guidelines

NICE (UK), British Society of Gastroenterology, American Gastroenterological Association. Cerebra defaults to guideline-consistent guidance.

licensed
GutIQ internal
15 years of de-identified clinical observations

Across 25,000+ consultations: symptom progressions, what worked, what didn't, intervention sequences. The pattern library that gives Cerebra its bedside intuition. Every record is stripped of names, emails, phone numbers, dates of birth, and free-text identifiers before any model sees it.

proprietary

Last full retraining: 2026-04 · Next scheduled refresh: 2026-08 · Continuous reference-document index updated weekly

04 · Privacy & Consent

Your record is yours.
You can opt out at any time.

We treat training data the way a regulated lab treats patient samples: identifiers stripped at source, access logged, and the right to withdraw is yours unconditionally.

Manage training-data preferences
No personally identifiable information ever reaches the model

Name, email, phone, address, date of birth, and any free-text identifier are removed by a dedicated pipeline before training data is staged.

Opt out of training. Keep your account.

A single toggle in your settings excludes all of your data — past, present, and future — from any further retraining. Your reports, your plan, and the chat itself continue working unchanged.

Chats stay on-platform

Cerebra runs on infrastructure we control. Conversations are not shared with third-party LLM providers and are never used for advertising.

Hard delete on request

Delete your account and we expunge associated records from the training pool. The next retrain re-runs without them.

05 · The Interface

Ask a careful question.
Get a careful answer.

A sample conversation. Notice the routed-experts strip up top, the citations under each answer, and the medical-only guardrail at the bottom of the composer.

C
CEREBRA
medical assistant · v2026.06
e_01 functional-gi e_03 nutrition e_07 trajectory +8 standby
A

I've been waking up with bloating for three weeks. It's worst after dinner, and a glass of wine seems to make it worse the next morning. Anything I should look at first?

C

The pattern you're describing — overnight accumulation that's worse after dinner and amplified by alcohol — is most consistent with delayed gastric emptying or fermentation in the distal small bowel. A few practical first steps that are well-supported in the literature:

1. Shift dinner earlier. A 12-hour overnight fast (e.g. 7pm → 7am) gives the migrating motor complex time to sweep the small bowel. This alone resolves overnight bloating in ~40% of cases.

2. Cut alcohol for 10 days. Alcohol delays gastric emptying and increases small-intestinal permeability; both will amplify next-morning bloating even when intake feels modest.

3. Watch for FODMAP density at dinner specifically. Onion, garlic, legumes, and wheat carry more fermentable load than people realise. We don't recommend a full exclusion yet — just track the meal context for a week.

If the bloating is still present after 10 days of (1) and (2), that changes the picture and we should look at SIBO testing as the next step.

Cited Cangemi 2019 · MMC & overnight motility Bode 2020 · alcohol & gastric emptying Halmos 2014 · FODMAP RCT NICE NG61 · IBS
Routed through e_01 functional-gi + e_03 nutrition · low-confidence on alcohol dose-response below 2 units/wk
A

Should I just start taking a probiotic?

C

Not yet — and probably not the way the supermarket aisle implies. Three things to know:

First, generic multi-strain probiotics show essentially no effect on overnight bloating in randomised trials. The strains that have evidence (Bifidobacterium infantis 35624, Saccharomyces boulardii for post-antibiotic recovery) are specific.

Second, starting a probiotic now would confound the 10-day dinner-timing trial. You wouldn't know which change moved the needle.

Third, if your bloating turns out to be SIBO-pattern, certain probiotics can worsen it temporarily by adding to the small-bowel fermentable load.

So: hold off, finish the trial above, and if symptoms persist we'll look at a targeted strain — not a generic blend.

Cited Ford 2018 · probiotics in IBS meta-analysis Rao 2018 · SIBO & probiotics caution
Routed through e_02 microbiome + e_04 pharmacology
C
Medical-only mode · off-topic queries are politely declined preview — composer disabled

Cerebra is for information and triage; it is not a substitute for an in-person clinician. Anything that suggests urgent harm is routed straight to a human consultant.

06 · Scope

What Cerebra will and won't do.

In scope
  • Interpret symptoms, food triggers, supplement choices.
  • Read your daily log series for trend signal.
  • Explain lab results in plain English with primary citations.
  • Suggest questions to bring to your consultation.
  • Compare evidence on common interventions (probiotics, fibre, FODMAP, etc.).
  • Triage red flags and recommend escalation when warranted.
Out of scope
  • × Non-medical chat. Cerebra politely redirects.
  • × Definitive diagnoses. Cerebra reasons; clinicians decide.
  • × Prescriptions or dose adjustments without consultant approval.
  • × Mental-health crisis management. Escalates to human + signposts crisis lines.
  • × Areas outside its training: cardiology, oncology specifics, surgical decision-making.
  • × Anything advisable only with imaging in hand.
07 · For the technically curious

Under the hood.

Base architecture
Open-weights foundation model, sparse mixture-of-experts head, top-2 routing.
Expert count
11 domain experts + 1 router. Two experts active per token on average.
Fine-tuning
LoRA on each expert against its domain corpus; SFT + DPO on a 60k-question clinical Q&A set for the router.
Grounding
Retrieval over a vector index of 9.2M abstracts. Every answer includes its top-k citations and the routed experts.
Safety layers
Red-flag classifier (always on), topic classifier (medical-only), self-consistency check across experts, abstention when confidence is below threshold.
Hosting
Dedicated inference on infrastructure we operate. No third-party LLM provider sees your conversation.
Evaluation
Quarterly benchmark on MedQA, PubMedQA, MMLU-Med, plus a held-out 1,000-question GI panel reviewed by our consultant board.
FAQ

Common questions.

Is Cerebra ChatGPT under the hood?
No. Cerebra runs on open-weights foundation models we fine-tune ourselves on medical data. We do not send your conversation to OpenAI, Anthropic, or any third-party LLM provider — the entire stack runs on infrastructure we operate.
How is this different from asking a general LLM?
A general LLM is a single model trained on the whole internet. Cerebra is a router plus eleven specialist experts, each fine-tuned only on its own medical sub-domain. The router activates the right experts per question, and the answer is grounded in a citation index — every claim links back to a paper or guideline.
What happens if I ask Cerebra something non-medical?
A topic classifier sits in front of the model. Non-medical queries get a short, polite redirect — Cerebra will tell you it's a medical-only assistant and suggest where else to ask. This is by design: a focused tool is a more useful tool.
Can I really opt out of training and still use Cerebra?
Yes — one toggle, no friction. Your existing reports, plan, and chat history all continue to work. The toggle simply marks your record as ineligible for inclusion in any future retraining pool. Already-deployed model weights cannot be selectively un-learned, but they were trained without your PII to begin with, and your data is excluded from every subsequent retrain.
Will Cerebra ever replace a doctor?
No — and we'd be suspicious of any AI company that promises it will. Cerebra is built to make the time between consultations more productive: better-informed questions going in, clearer follow-through coming out. The judgment call is still a clinician's.
How often does Cerebra get retrained?
Major retraining is quarterly. The retrieval index (the citation library Cerebra grounds answers in) is updated weekly so that newly published evidence appears in answers without a full retrain.

Cerebra ships inside every PURO™ and CORE™ programme.

And a free preview ships with every account. Take the quiz, get your gut profile, and the chat opens with the first three questions worth asking already drafted.

Cerebra is a registered research product of GutIQ. Not a medical device.