The front end you liked, with the backend actually built

Turn a raw DNA file into a fitness blueprint a coach can actually use.

A genetics module built to drop straight into your stack. Real interpretation, every result traceable to a source, and the human coach kept right in the loop.

Built native to Lovable so you can keep adding, modifying, and growing it yourself.
🧬
Your Blueprint
Sample report · M, 34
⚙️
Genetic readiness index
82/100
Strong power and recovery genetics. Build around strength blocks, manage caffeine timing, watch vitamin D.
Power dominant Fast recovery Slow caffeine
Top signals for you
⚡ Power vs endurance
ACTN3 · RR
The RR genotype is associated with the fast twitch muscle profile common in power and sprint athletes.
Evidence Well establishedGWAS Catalog · SNPedia
☕ Caffeine metabolism
CYP1A2 · AC
Slower caffeine clearance. Keep doses earlier in the day to protect sleep and recovery.
Evidence Well establishedPharmGKB
🦴 Injury and tendon load
COL5A1 · CT
Moderate connective tissue consideration. Ramp volume gradually and prioritize eccentric work.
Evidence EmergingPeer reviewed
Training genetics
⚡ Muscle fiber type
ACTN3 · RR
Lean toward strength and power blocks. You respond well to heavy, lower rep work and explosive movement.
Evidence Well establishedGWAS Catalog
👤 Coach layer

Your coach sees this and programs a 4 week strength base before any conditioning push. The app suggests, the human decides.

🫀 VO2 max trainability
PPARGC1A · GS
Above average aerobic response to training. Intervals will move the needle faster than for most.
Evidence EmergingPeer reviewed
💪 Strength response
ACE · DD
DD genotype is linked to power and strength adaptation. Favor progressive overload over high volume.
Evidence Well establishedSNPedia
Nutrition genetics
🥛 Lactose tolerance
MCM6 · GG
Likely reduced lactase persistence. Watch dairy heavy meals around training.
Evidence Well establishedClinVar
☀️ Vitamin D processing
GC · rs2282679 · GT
Predisposed to lower circulating vitamin D. A test and a coach guided protocol is worth it.
Evidence Well establishedGWAS Catalog
👤 Coach layer

Flagged for your coach to review with a real blood panel. The app never diagnoses. It hands the human a head start.

🧂 Salt sensitivity
AGT · MT
Moderate sensitivity. Mind sodium on high sweat training days.
Evidence EmergingPeer reviewed
Recovery and sleep
🔁 Recovery speed
IL6 · GG
Favorable inflammatory recovery profile. You can handle a higher training frequency than average.
Evidence EmergingPeer reviewed
🌙 Sleep chronotype
PER3 · long
Slight evening lean. Push hard sessions to later in the day when you can.
Evidence EmergingSNPedia
⬆
Drop your raw DNA file 23andMe, AncestryDNA, or any raw export. Parsed on a secure backend, never sold.
1
Parse the raw file
rsid, genotype, chromosome, position read into a clean table.
2
Match against the curated database
Deterministic lookup. Your genotype is mapped to known, sourced associations.
3
Explain in plain language
Claude writes the human friendly summary from the retrieved facts only.
4
Hand it to the coach
Anything clinical is flagged for a real human, never auto diagnosed.
🎯

Every result is traceable

No insight appears without a genotype and a source behind it. If we cannot ground it, it does not ship to the user.

🤝

The human stays in

The engine prepares, the coach decides. Exactly the model you described, AI that strengthens training instead of replacing the trainer.

🧩

Yours to grow

Built as clean components and a typed backend so it lives inside Lovable and you can keep shipping on it without me.

The part you have been fighting

Why this one does not hallucinate

The trick is to never let the model invent the science. Retrieval is deterministic and sourced. The model only does the last mile, turning facts you already trust into language a member understands.

Step 1

Raw DNA file

23andMe or Ancestry export. Hundreds of thousands of rows.

▶
Step 2

Deterministic parse

Code, not a model, extracts rsid and genotype into a table.

▶
Step 3

Curated lookup

Genotype joined to a sourced knowledge base. A fixed table, zero guessing.

No hallucination here▶
Step 4

Grounded write up

Claude explains only the retrieved rows. It cannot add facts.

▶
Step 5

Coach review

Clinical flags routed to a human before the member sees them.

What kept breaking, and the fix

  • ✓
    The model was asked to know genetics. Now it is only asked to rephrase facts we already retrieved.
  • ✓
    Genotype to trait is a database join with citations, so the same input always returns the same sourced result.
  • ✓
    Each insight carries an evidence tier and a source, so weak associations are labeled, not hidden.
  • ✓
    If a marker is not in the curated set, the app says so instead of improvising.

Suggested stack

Front endLovable, native components
App generationLovable on Gemini
Interpretation engineClaude, grounded only
Data and authSupabase + edge functions
Knowledge baseCurated SNP set, versioned
OwnershipAll yours, inside your project

Stays native to Lovable

You said keep it inside Lovable so you can add, modify, and grow it. That is the whole design. The Claude interpretation runs as a clean backend function Lovable calls. Nothing locks you out of your own product.

This is a concept, not the finished product. The real one is built around your data and your brand.

Pick up where you left off in Lovable, or rebuild the backend clean. Either way, you keep the front end you liked and finally get an interpretation engine you can trust.

Wellness and performance context only. Not a medical device and not a diagnosis. Genetic associations vary by population and evidence strength, which is why every insight shows its tier and a real coach reviews anything clinical. Sample genotypes shown are illustrative for this concept.