The strongest LLMs couldn't read my DNA file, so I built the layer they were missing

Hi Indiehackers, I'm Ida and I'm a solo founder of Helisoma

Want to add some context: I wear Whoop, I've watched the Huberman Lab podcast for years and worked through most of the protocols, I love cold plunges, saunas and zone-2 training. Anything that helps me understand my own body a bit better, I'm in. And of course I did a DNA test years ago. I got the ancestry pie chart, a few findings written in medical terminology I didn't understand a word of, and then I forgot about the whole thing.

And of course I was interested in using an LLM as a health coach, so at some point I ran that same old raw DNA file through the strongest models available, at maximum reasoning effort, and asked each of them the same questions about coffee, sleep and training. Every answer came back confident, specific and well written. The problem is they also contradicted each other. Same file, same marker, opposite verdicts. For example - one model told me I was a slow caffeine metaboliser and should stop drinking coffee after midday, another read the same genotype and said I clear it fast. 

That is what sent me to SNPedia to look up the most popular SNPs one by one by hand and work out which model was right. Checking them one at a time is also how I found out what was actually going wrong and it was something different every time. "Sorry, I had the orientation backwards" "Sorry, I made that rsID up”, "Yes, you're absolutely right, that's a palindromic SNP”.  That's when it clicked: a language model simply cannot carry that much arithmetic across that much data (average file has around 600,000 rows).

But I couldn't drop the idea of talking to an AI about this, because that part had genuinely been good. Nothing else will explain what "linkage disequilibrium" or "odds ratio 1.2" means for you personally, in words you can use and then answer the same question a third time without making you feel slow. The problem was never the LLM. It was that I was handing it a file it couldn't read properly.

So I decided to build the database of common SNPs and traits myself. A couple of years ago this would probably have taken me a year. God bless living in the age of AI, which helped me work through a pile of problems I had no background in: which allele a study measured and which strand your file wrote it on, hemizygous X calls padded to two letters so one copy looks like two, effect sizes published in whatever unit each paper felt like and reference alleles that turn out to be the rare one.

It also taught me how to find and actually read the papers. That's how I learned that most popular services are built on studies with very small samples, which makes their conclusions closer to horoscopes than to anything medical. Study quality mattered to me, so every finding in my report states how strong the effect actually is.

Now the scope, because this is where I expect the pushback. There are no disease risk scores in the report - no cancer, no Alzheimer's, no diabetes risk, no thrombophilia, no BRCA. I'm a software engineer and definitely not in a position to publish that kind of things, but if someone with a medical background wants to join the project as a co-founder, I'd be glad to work on that part together (text me ida@helisoma.com).

What reaches a report is the everyday stuff: caffeine, sleep timing, how you handle training loads, nutrient absorption, etc. Behind it is an evidence base I built from the GWAS Catalog, PharmGKB, ClinVar and CPIC: 37,742 association rows across 885 markers, collapsed to 1,145 entries. A marker carries forty published associations on average, from studies of wildly different quality, so most of the work was deciding which one speaks for it. Each entry is oriented and weighted by published effect size and links to its paper. Please keep in mind that the studies behind all of this were run mostly on people of European ancestry. That's the world we live in, unfortunately.

Mechanically: the raw file is parsed in your browser and never leaves your device (you can check the network). Out of roughly 600-700k markers, depending on the chip, only the ~600 the report needs get sent. What comes back is one report you can read as a PDF or attach to your own Claude, ChatGPT, Gemini or Whoop as an MCP connector or a file bundle (there's no chat on my side, use it with your own AI).

Random facts I found about myself, since that's the real test:

• Coffee. I had banned myself from drinking it after lunch for years. But my CYP1A2 genotype at rs762551 is A/A, the fast-metaboliser version, so in theory caffeine clears quickly for me. I tried drinking it all day and found out that in my case it genuinely doesn't touch my sleep.

• Tendons and ligaments. One marker leans toward more strain there, which matches my life very well. So pushing hard in a new sport and trying to keep up with everyone else is exactly what I shouldn't do.

• Alcohol. The pull to drink builds a little more easily than average for me. I decided it's simpler to skip it altogether than to spend energy holding myself in check.

• A leaning toward low iron, read across seven markers. Iron, ferritin and anaemia are a common problem for women and I'm no exception. I went and got tested, it was confirmed, and I'm on iron now.

My favourite part was pairing it with the Whoop AI chat. I already log a lot in the Whoop journal, so it knows my habits and my recovery data, and connecting all of that with the DNA test worths the whole project

The app: https://helisoma.com/, $49 once, no subscription, the preview insights are free. It stays cheap because the expensive part, the conversation, runs on the AI you already pay for. The sample pdf report is on the landing page (with my own data). I'd love any feedback.