System Architecture · v1.0

Project Rosie

Personalized cancer vaccine pipeline · software intelligence layer

Vertex AICloud RunSupabasegemma-4-26b-a4b-it-maasGemma4Good

Watch a VCF flow end-to-end

System Layers

Browser / Vet Clinic

Next.js 16

App Router · Deployed on Vercel

Upload VCFPipeline TimelineCase DashboardReport ViewerSensitivity SlidersChat Widget

Supabase Auth · email/password · RLS

Platform

API + Storage

Next.js API Routes

POST /api/casesPOST /pipeline/callbackPOST /api/gemma/*

Google Cloud Storage

VCF uploads (resumable) · pipeline outputs + PNG charts · 30-day lifecycle · us-central1

Supabase

PostgreSQL · Realtime subscriptions · RLS
Receives per-stage callbacks → pushes to browser

Cloud Run Job

Pipeline

4 stages · 5 status callbacks

running10–15 min

Neoantigen Prediction

VEP v115 · CanFam4 ROS_Cfam_1.0 · pVACseq + NetMHCpan + IEDB API · DLA-8850101 + DLA-8850801 · 8–11mer peptides

scoring1–3 min

Candidate Ranking

IC50 < 500 nM filter · VAF-weighted composite score · Top 20 candidates → JSON · auditable ~50-line script

reporting2–5 min

Report Generation✦ Gemma

Binding affinity + mutation landscape charts → GCS · Gemma 4 reads JSON + 2 PNGs multimodally → 400–600 word species-aware clinical report

designing1–2 min

mRNA Vaccine Design

Biopython · canine codon table · 5′UTR + Kozak + AAY linkers + poly-A(60) · Synthesis spec via Jinja template (not LLM-generated)

4 vCPU / 8 GiB / 3h timeout · us-central1 · service-account auth

Gemma 4 · Vertex AI Global Endpoint

AI Roles

256K context · multimodal vision · 4 functions

Role 1 · Pre-flight · Proactive

VCF Advisor

Browser sends VCF structural facts → 0–3 issues (info/warning/critical) as strict JSON · zod safe-parse · never blocks submit

Role 2 · Post-pipeline · Vision

Multimodal Report Writer

Reads candidates JSON + binding-affinity bar chart + mutation-landscape pie (both PNGs as images, not OCR) · species-aware prompt (DLA/HLA/FLA)

Role 3 · Interactive · On demand

Sensitivity Narrator

IC50 + VAF sliders re-rank client-side instantly (no Gemma cost per tick) · on explicit click, Gemma narrates the kept/dropped tradeoff in plain English

Role 4 · Reactive · Conversational

Case Assistant

Floating chat widget · full case context in system prompt every message · plain-language Q&A for veterinary oncologists

Pipeline Outputs

Ranked Candidates

Neoantigen List

Top 20 · JSON · IC50 + VAF composite score

MPMCEFDMVK · IC50 128 nM · DLA-8850801 · PIK3CA V125M

Gemma Output

Clinical Report

400–600 words · plain language · species-aware · multimodal input (true vision, not OCR)

Sequence File

mRNA Vaccine

FASTA · ~251 nt · 54.8% GC (CDS) · canine codon table

5′UTR + Kozak + CDS + AAY linkers + 3′UTR + poly-A(60)

Deterministic

Synthesis Spec

PDF · Jinja2 template — not LLM-generated · CleanCap® AG cap1 · 100% m1Ψ · SM-102 LNP · RIN ≥ 8.0

"AI for interpretation, deterministic code for compliance."

rosie.kiraklabs.com · ~1 hour end-to-end · ~$15/case · 191,645 epitopes → 1,648 after IC50 filter → top 20 ranked

Gemma4GoodVeterinary oncologyPersonalized medicine