How the engine works.
VitSync builds your plan in two layers. An on-device rules engine always runs and scores every ingredient against your data. If you agree to it, an AI layer, Anthropic’s Claude running on our server, reads the same signals and writes your plan and its explanations. The AI layer works only inside what the safety gate allows. Every recommendation traces back to specific data points and specific studies.
1. Signal collection
The engine knows nearly 40 signals, and each ingredient reads only the ones research links to it. They come from seven sources:
- Apple Health: Sleep duration and quality, HRV, active minutes, workouts, body weight (for dose bands), menstrual flow if you allow it, and nutrition if you log it in apps like MyFitnessPal.
- Diet and drinks: Fruit and veg, leafy greens, oily fish, meat/eggs/dairy servings, whether you are vegan or vegetarian, cups of tea and coffee a day, and alcohol a week.
- Self-reported: Energy, stress, menstrual flow, training hours, outdoor hours, and sleep when Apple Health has none.
- Blood results: Vitamin D, magnesium, folate, B12, ferritin, testosterone, and for women the hormones oestradiol, progesterone, FSH and AMH (all optional, entered in-app).
- Health and life stage: Conditions such as PCOS/PMOS or an autoimmune thyroid condition, menopause stage, trying to conceive, statins or metformin, and age.
- Chat context: Health events (sick, travelling, jet lag), lifestyle facts (vegan, trains 6x/week), upcoming events.
- Environment: UV index and daylight hours via WeatherKit, adjusted for your location and season.
Apple Health data is processed on-device. Only computed summary signals are transmitted during plan generation. Raw HealthKit samples never leave your phone.
2. Sigmoid scoring
Each signal is converted into a 0 to 1 need score using sigmoid curves. These are not linear clamps. The curve is calibrated against your personal baselines, your 30-day HRV and sleep averages, refreshed monthly.
This means the engine responds to relative changes in your data, not absolute thresholds. A HRV of 45ms might be fine for you but concerning for someone whose baseline is 70ms.
For each ingredient, need scores are multiplied by evidence-weighted signal strengths (each signal carries a weight reflecting how directly it relates to the ingredient) and normalised into a single combined score between 0 and 1.
If the combined score reaches 0.25, the ingredient is included in your plan (0.30 for narrower-evidence ingredients that need a stronger signal fit). If it drops below 0.14, it is removed. The gap between these two thresholds (hysteresis) prevents supplements bouncing in and out on a single bad night.
Some answers can only ever raise a score. Tea and coffee reduce how much iron you absorb from a meal, so more cups can nudge iron up and never down. Alcohol works the same way for the ingredients it affects. A light answer never scores lower than no answer at all. Hormone markers can support an ingredient but never add one on their own.
3. Evidence grading
Every signal-to-ingredient link is graded by evidence strength:
- RCT-grade: Direct randomised controlled trial or official body guidance (NHS, EFSA, NIH) linking the signal to the ingredient benefit.
- Plausible: Mechanistically sound link supported by observational studies, cohort data, or established biochemistry, but lacking a direct RCT for this specific use case.
The fraction of RCT-grade signals contributing to an ingredient's score directly affects its confidence tier. This is not just about how high the score is, but how trustworthy the evidence behind that score is.
4. Confidence tiers
Every recommendation is assigned one of three confidence levels:
Blood-verified
Your own blood result for that ingredient’s main marker points to need at 0.55 or more, the reading sits below the laboratory’s low limit for that marker, the result is under 180 days old, and at least 60% of the contributing evidence is RCT-grade.
Recommended
At least 3 signals agree, the score is 0.34 or above, and at least 50% of the contributing evidence is RCT-grade.
Consider
The recommendation is supported by the data but with fewer contributing signals or a higher proportion of plausible-grade evidence.
The badge on every ingredient card in the app reflects this tier. You can always see why something is recommended and how confident the engine is.
5. Safety gate
Before scoring even begins, a safety gate evaluates your declared medications, conditions, and allergies. Any ingredient with a known interaction is blocked or flagged before the engine scores it.
This is structurally separate from the scoring logic. A high need score cannot override a safety block. The gate checks for:
- Pregnancy, breastfeeding, trying to conceive
- Oral contraceptive use
- Under 18
- Anticoagulants (warfarin, heparin, DOACs)
- Thyroid medication
- Statins
- ADHD medication
- SSRIs and SNRIs
- Kidney disease
- Fish and shellfish allergies
- Immunosuppression
- Smoking or asbestos exposure (blocks beta-carotene)
- Diabetes medication
- History of a hormone-sensitive condition
- Vegan diet (triggers B12 reinforcement)
- GI conditions (Crohn's, IBS, coeliac)
- High ferritin (blocks iron)
The AI layer is bound by the same gate. The app runs the safety gate over every plan the AI layer returns and drops any row it has blocked, so the AI cannot add back something the gate removed. If the AI layer is unavailable, or you chose not to use it, your plan comes from the on-device engine alone.
Iron is always review-gated. The engine never auto-recommends iron without flagging it for GP review first, regardless of your ferritin levels or signals.
6. Gender-specific adjustments
The evidence base for several ingredients differs significantly between men and women. The engine applies sex-specific multipliers based on published research:
- Women: Menstrual flow influences iron signals. Folate weighting is increased. Cycle phase (if tracked) adjusts myo-inositol and vitex scoring. Post-menopausal status shifts K2 and D3 weightings for bone health.
- Men: Testosterone-linked supplements (creatine, zinc) use a different evidence weighting. Prostate-relevant interactions are checked.
7. Interaction effects
The engine does not score each ingredient in isolation. Compounding signals are detected:
- When sleep, HRV, and stress are all poor simultaneously, those shared signals compound. This lifts the recovery-relevant ingredients (magnesium, ashwagandha, omega-3) together rather than scoring them independently.
- Training load uses the maximum of HealthKit data and your self-reported hours, so weekend warriors who train hard but irregularly are not undercounted.
- When both HealthKit nutrition data and manual diet answers exist, the engine blends them at 70% real data / 30% manual to smooth over logging gaps without discarding your answers entirely.
- Chat-derived health events (sick, travelling, jet lag) apply a temporary 7-day signal adjustment rather than a permanent profile change.
8. Weekly or monthly adaptation
Your plan is rebuilt from fresh data every week, or every month if you choose that in Profile. This is not a static recommendation that was set once during onboarding. At each refresh the engine re-evaluates all 27 ingredients against your latest signals.
The hysteresis thresholds (0.25 to add, 0.14 to remove) mean the plan only changes when patterns hold. A single night of poor sleep will not add magnesium. A week of poor sleep can.
Each refresh, the plan shows what changed and why: what was added, what was removed, and the signals that drove those changes.
9. Evidence monitoring
The evidence base is reviewed monthly against the latest published research from PubMed, NHS, and EFSA. A scheduled review surfaces new systematic reviews, meta-analyses, safety alerts, and guideline changes.
Changes to recommendation weightings or claims are never made automatically. The monthly review produces a report that is assessed manually before any updates reach the app.
10. What the engine deliberately doesn't do
It doesn't diagnose. It doesn't treat. It doesn't prescribe. It doesn't replace your doctor, your pharmacist, or your dietitian. If your data points to a serious symptom in chat, the assistant directs you to your GP, 111, or 999.
It doesn't sell you supplements. The Buy on Amazon links earn us a small affiliate commission but the engine doesn't know about that. Recommendation logic is fully separate from product selection.
Questions
If you want to know why a specific ingredient is in your plan, tap any card in the app and the explanation expands. If you want to know more about the engine itself, email us at hello@vitsync.com.