NeuraLog AI

AI Research Infrastructure for Clinical Trials

1 follower

AI Research Infrastructure for Clinical Trials

1 follower

Visit website

Automation tools

•

Databases and backend frameworks

•

Real-time collaboration infra

Researchers spend 15–40 minutes per patient visit manually pulling clinical scale scores from notes. NeuraLog AI does it in seconds and shows you the exact sentence every value came from. What's different: It refuses. When documentation is ambiguous or conflicting, it returns null and flags it instead of inventing a plausible number. Built by a physician, so it knows HAM-D-17 and HAM-D-21 aren't the same scale. Every scale citation verified against PubMed.

Payment Required

Launch tags:SaaS•Artificial Intelligence•Health

Launch Team / Built With

AWS Activate CreditsCredits that fuel your startup’s next milestone

Promoted

Vercel

Vercel Day

Which Vercel product did you use? What new capability did it give your team, and how did that shape the product you launched?

Maker

Next.js App Router on Vercel - specifically server-less API routes. The capability that mattered: a HIPAA-adjacent product needs PHI redaction to run server-side, before any text reaches an external model. Vercel's server-less functions gave me that boundary with no infrastructure to manage. Every uploaded clinical note passes through an 18-category HIPAA Safe Harbour redaction pass in a server-less route before a single character reaches the Claude API. My API credentials never touch the client bundle. How it shaped the product: I'm a physician, not an engineer, building this solo while completing medical board exams. If I'd had to stand up and secure my own backend, this wouldn't exist. Preview deployments meant I could ship a fix, verify it live, and roll back safely without a DevOps person. Thirty-one routes in production, deployed from my laptop, built around clinical rotations. Vercel is the reason a solo clinician could ship production healthcare software at all.

Report

2mo ago

Maker

📌

Hi Product Hunt! I'm Dr. Rameen Rao, a physician and dementia researcher. I built NeuraLog AI because I kept watching the same thing happen in research labs: someone opens a patient chart, reads a note, finds an NIHSS or MMSE score, types it into a spreadsheet, closes the chart, opens the next one. Fifteen to forty minutes per visit. For a 200-patient study across three timepoints, that's hundreds of hours before anyone gets to do actual science. The obvious answer is "just use AI." I tried. General models will confidently hand you a number that isn't in the note. They'll conflate HAM-D-17 and HAM-D-21, which have different denominators. They'll give BDI-II the wrong severity bands. And there's no way to check - no citation, no audit trail, nothing. In clinical research, a wrong number that enters a published dataset is far worse than a missing one. So I built this around a principle I ended up calling "null over wrong": if the documentation doesn't support a score, the system returns nothing and flags it. Click any extracted value and it highlights the exact sentence it came from. If it can't verify the source, it says so rather than guessing. The part that surprised me most: building refusal is harder than building extraction. I audited my own product and found places where it was inferring scores from narrative descriptions instead of refusing, exactly the failure mode I built it to prevent. Fixing those became the roadmap. I also ran every clinical scale citation against PubMed and found a 16% error rate, including two citations that didn't exist. All corrected, and nothing goes in now without independent verification. Happy to talk about any of it: the clinical logic, the refusal architecture, the de-identification layer, or what it's like building software as a physician. If you do clinical research, what's your current abstraction workflow? I'd genuinely like to know what I'm missing.

Report

2mo ago

Pros

Cons

Reviews