Conceiva

Inspiration

One of our team members ran into how hard it is to enter the fertility industry — and how wasteful the pipeline can be. Digging into sperm-bank operations, a stark number kept coming up: about 90% of sperm samples are rejected and never make it into usable inventory. That is not only a logistics problem for banks; it is a bottleneck for the families waiting on donors.

At the same time, couples who do find candidates still face a second gap. Extended carrier screening is long. Shared recessive risks are easy to miss in a spreadsheet. Genetic counselors often spend the first half of a visit reconstructing which donors are even in play. Couples, banks, and counselors sit in separate systems with no shared shortlist or briefing.

Conceiva started from those two pressures: help banks surface sample quality earlier (including motility from real microscopy video), and help couples move from “interesting profile” to “counselor-ready shortlist” without pretending the product is a diagnosis.

What it does

Conceiva is decision support for family building with genetics in the loop.

Couples enter carrier and preference context, swipe through donors (baby-photo–first cards), see hard carrier conflicts before attachment, shortlist candidates, and book a genetic counselor who can see that shortlist.
Sperm banks manage donor catalog records and can run a motility analysis on an uploaded microscopy clip (YOLO tracking → progressive / non-progressive / immotile metrics vs WHO reference floors).
Genetic counselors open visits already oriented to the couple’s shortlist, with AI-assisted plain-language summaries of carrier complications and gaps — so the session starts deciding, not catching up.

It is not a lab, a diagnosis, or a substitute for a genetic counselor. It is an intermediary on records the users supply: match → shortlist → prepared counseling visit.

How we built it

Web app: FastAPI + Jinja + SQLite, session auth, role-gated couple / bank / counselor flows.
Matching: Carrier hard-stops, blood-type flags, preference soft weights, swipe-style match UI over a seeded demo catalog.
Counseling path: Shortlist, availability, booking, LLM-assisted counselor reports grounded in case data.
Motility service: Separate FastAPI service (Ultralytics YOLO + ByteTrack), H.264 annotated clips via ffmpeg, metrics written back to the donor.
Deploy: Vercel for the site; a cheap single EC2 host for multi-minute CPU analyze (serverless cannot run it). Stripe optional for unlocks.

Challenges we ran into

Motility cannot live on Vercel (torch + 1–2 min jobs) → split deploy and a public service URL.
First builds pulled ~2.5 GB CUDA wheels onto a CPU box → switched to CPU torch + S3→EC2 bootstrap on t3.small.
Missing production ffmpeg broke annotated clips until we installed it (plus an imageio-ffmpeg fallback).
Vercel /tmp SQLite reseeds made consent “forget” and loop → session-backed consent + pre-consented demos.
Parallel branches briefly overwrote a UI redesign on prod until we merged main back in.

Accomplishments that we're proud of

Full path: bank catalog → couple match/shortlist → counselor visit with AI briefing.
A real motility pipeline reachable from the shared demo (not a fake precomputed upload).
Product copy that stays clinically humble: decision support, hard recessive stops, WHO comparison without claiming diagnosis.
A weekend-viable deploy: Vercel + one cheap EC2 for the heavy model.

What we learned

Fertility products only work when couples, banks, and counselors share the same shortlist context.
Minute-scale inference forces architecture choices; you cannot pretend the website host is a GPU lab.
Serverless demos need designs for resetting state, not a durable laptop disk.
Small ops details (ffmpeg, CPU vs CUDA, no-ALB EC2) decide whether the live demo survives judges.

What's next for Conceiva

Durable DB (not /tmp SQLite) for multi-user pilots.
Calibrated motility and deeper bank integrations.
Stronger counselor tooling: structured panel imports, clearer gap reports, cited visit notes.
Privacy/compliance hardening beyond synthetic demo genomes.
Stay true to the spark: less wasted sample inventory upstream, better-prepared counseling downstream.