2 of 10

02 · SECURITY & PRIVACY

ONKOS

Oncologists are buried in patient data,

and patients are the ones waiting.

Every new patient shows up with hundreds of pages of PDFs, scans, genomic reports, and handwritten notes. Oncologists have to read through it all by hand, then cross-check it against the eligibility criteria for hundreds of open trials. It takes days. Patients wait. Trials that could have helped them quietly miss their enrollment windows.

MATCHING TAKES DAYS

75%

of clinical trials don't hit their original enrollment timelines.

Cutting Edge Information; Tufts CSDD

200+

pages in the average new patient's chart.

Internal EHR audits, NEJM Catalyst, 2018

40 Days

to manually screen one patient against a trial shortlist.

JCO Clinical Cancer Informatics, 2021

AND THE DATA KEEPS PILING UP

59%

of U.S. oncologists report burnout symptoms, the highest of any medical specialty.

Hlubocky et al., JCO Oncology Practice, 2023

2 hrs

on paperwork and EHR work for every hour spent with a patient.

Sinsky et al., Annals of Internal Medicine, 2016

4,000+

data points in a typical cancer patient's record, and most of it isn't searchable.

IBM Watson Health analysis, 2018

01 / 10

HackPrinceton 2026 · Regeneron Track

3 of 10

02 · SECURITY & PRIVACY

ONKOS

02 / 10

HackPrinceton 2026 · Regeneron Track

HIPAA-aligned by design.

No real patient data enters the system. Logs are scrubbed before they touch disk, endpoints can be token-gated, and case state never leaves memory.

SYNTHETIC DATA

No real patient records, ever.

Every demo case is generated from scratch.

No PHI, no IRB, no burden.

REDACTED LOGS

Regex scrubber runs on every log line.

Catches SSN, DOB, phone, email, MRN.

On by default (NEOVAX_LOG_REDACTION).

TOKEN-GATED API

Every /api/* can require a bearer token.

Constant-time comparison (secrets.compare_digest).

Flip on with NEOVAX_API_TOKEN.

EPHEMERAL CASES

Cases live in memory for one session.

Nothing persists to disk.

A server restart drops every case.

ALIGNED, NOT CERTIFIED

Full HIPAA compliance needs a BAA with every vendor (K2-Think, Kimi, HeyGen, Google Maps). This is the foundation, not the finish line.

4 of 10

03 · ARCHITECTURE

ONKOS

03 / 10

HackPrinceton 2026 · Regeneron Track

How the pieces fit together.

A Next.js frontend talks to a FastAPI backend over server-sent events. Model output is pushed to the UI as soon as it arrives.

FRONTEND

Next.js 15 · Tailwind

Upload flow and live cockpit

Clinician tabs: Overview, Plan,�Trials, Documents, Clinical

Patient sidebar with spoken avatar

Tab state lives in the URL

BACKEND

FastAPI · SSE · Python 3.11

Orchestrator walks the pipeline

Case store and event bus in memory

PDF report generator

Google Maps geocoding

MODELS & DATA

K2-Think · Kimi · HeyGen

K2-Think: reasons through�guidelines, shows its work

Kimi: handles chat with tools

HeyGen: gives the avatar a voice

PubMed index for citations

Regeneron trial registry

5 of 10

04 · SYNTHETIC DATA

ONKOS

04 / 10

HackPrinceton 2026 · Regeneron Track

FAX REFERRAL

Scanned outside referral with handwriting and stamps, simulating real-world document variety.

LAB CHART

Comprehensive CBC and CMP values tracked over time to demonstrate longitudinal data handling.

PATHOLOGY PDF

Digital biopsy report featuring IHC and a full molecular panel for advanced diagnostic simulation.

Built on realistic, synthetic data bundles.

Every demo case is synthetic: pathology PDFs, faxed referrals, lab charts, and registration sheets. It mirrors the "mess" of a real intake process with zero risk to patient privacy.

6 of 10

05 · REGENERON RELEVANCE

ONKOS

05 / 10

HackPrinceton 2026 · Regeneron Track

Built into the Regeneron oncology pipeline.

Trial matching runs on real eligibility rules from real Regeneron-sponsored studies. Not just a keyword search of ClinicalTrials.gov.

Regeneron trials�covered.

Biomarker rules�built in.

Eligibility checks�per trial.

3-way

Eligible, needs data,�or ineligible verdict.

KEY REGENERON ASSETS COVERED

Fianlimab + cemiplimab

LAG-3 with PD-1 in advanced melanoma.

Cemiplimab monotherapy

PD-1 across solid tumors.

Odronextamab

CD20×CD3 bispecific for B-cell lymphoma.

Linvoseltamab

BCMA×CD3 bispecific for multiple myeloma.

7 of 10

06 · HOW IT REASONS

ONKOS

06 / 10

HackPrinceton 2026 · Regeneron Track

The reasoning model drives it.

The NCCN treatment pathway is actually walked by a reasoning model, step by step. You watch it think in real time.

Live reasoning stream

K2-Think thinks out loud at each guideline step. The cockpit renders it as it happens.

Grounded in real papers

Each step pulls relevant PubMed citations, filtered to the patient's cancer type.

Works across cancer types

The walker picks the right branch based on the cancer type the extractor found. Nothing is hard-coded.

Follow-up chat

A second model handles questions after the run. Its tools only guide the UI; they never change the case.

The avatar

Kimi writes the answer. HeyGen gives it a voice. The avatar doesn't think on its own.

K2-THINK · LIVE

thinking... patient presents with Stage IIIC cutaneous melanoma, BRAF V600E positive, ECOG 1. primary decision: adjuvant vs. neoadjuvant systemic therapy.��per NCCN MEL-C: for resectable stage III with positive sentinel node, consider neoadjuvant pembrolizumab or nivolumab plus relatlimab (RELATIVITY-098).��flag: patient may also be eligible for a LAG-3 combo. cross-check Regeneron fianlimab + cemiplimab (NCT05352672)...

8 of 10

07 · DESIGN

ONKOS

07 / 10

HackPrinceton 2026 · Regeneron Track

One case. Two audiences.

The same case data powers two views. The avatar can send the user to the right tab when it needs to.

CLINICIAN COCKPIT

For the treating oncologist

Overview

Plan

Trials

Documents

Clinical

Intake and pathology at a glance

Guideline pathway with reasoning at each step

Ranked Regeneron trials with clear verdicts

One-click clinician PDF report

PATIENT SIDEBAR

For the person it's about

Diagnosis

Plan

Healing

Questions

Plain-English headlines, no jargon

The avatar reads answers aloud

Questions to ask the care team

Trial options explained in lay terms

9 of 10

08 · READINESS

ONKOS

08 / 10

HackPrinceton 2026 · Regeneron Track

Open-source, reproducible, ready to run.

MIT licensed. One command for the backend, one for the frontend. If a key is missing, that feature switches off instead of crashing.

SHIPPING TODAY

Pulls oncology fields out of PDFs

Guideline walker with live reasoning

Patient trial matching on a map

One-click clinician PDF report

Clinician and patient views

IN PROGRESS

Wire Kimi chat into the avatar

Capture more intake (ECOG, RECIST, prior tx)

Expand trial coverage

PRIVACY & DEPLOY

Synthetic data only. No real patient data.

Cases live in memory, one session at a time

Secrets in .env, a health endpoint shows setup

Vercel for the frontend, Render or Fly for the backend

1 of 10

2 of 10

3 of 10

4 of 10

5 of 10

6 of 10

7 of 10

8 of 10

9 of 10

10 of 10