PainFree

Inspiration

I am a solo non-developer creator. I don't write code for a living, but I know firsthand how exhausting it is to leave the hospital after a procedure with a multi-page itemized bill full of obscure codes, while dealing with severe post-op pain. Doctors ask patients to rate their pain from 1 to 10, which tells them nothing about how their actual daily mobility collapsed. Meanwhile, hospital billing offices expect regular people to decipher complex itemized charges on their own.

I wanted to build a practical tool that connects these two sides: capturing daily pain levels alongside objective daily activity drops, and using NVIDIA Nemotron models hosted on the Nebius Token Factory to audit photographed bills and draft clear, polite inquiry letters for hospital billing departments.

What it does

PainFree lets patients log their pain intensity (VAS 0–10) and mark precise anatomical pain points on interactive SVG body maps. It tracks daily activity metrics against the patient's personal baseline to highlight functional mobility drops, such as a 62% step reduction during acute flares.

When a patient uploads or photographs a hospital bill, the app runs an automated three-stage pipeline:

1. nvidia/Nemotron-3.5-Super-VL transcribes the photographed bill via Nebius Token Factory.
2. A deterministic server-side parser extracts line items, codes, and totals.
3. nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B cross-references the patient's condition and wearable telemetry with the billed charges.

The system generates a polite inquiry letter for the hospital billing office. Importantly, it separates in-hospital procedural drugs (such as Propofol and Rocuronium administered under clinical supervision) from discharge precautions, ensuring patients do not receive misleading home-care advice for surgical medications. Patients can also export a 1-page clinical summary report correlating 7-day pain levels with daily activity telemetry for doctor visits.

How I built it

As a non-developer creator, I built this entire project by pairing with autonomous AI coding agents:

• I directed Google Antigravity to architect the core pipeline, Python micro-server, and responsive TailwindCSS frontend.
• I tasked OpenAI Codex to automate browser validation and record our official high-definition walkthrough demo video via Playwright.
• Both Nemotron vision (Nemotron-3.5-Super-VL) and reasoning (Nemotron-3-Nano-30B-A3B) run live on high-throughput Nebius Token Factory endpoints.
• To keep the system trustworthy, all bill parsing happens deterministically on the server to prevent manipulated client payloads (server_reparsed), backed by strict JSON Schema guards.

Challenges I ran into

When you direct AI agents as a non-developer, the biggest danger is silent AI hallucination. In early tests with surgical bills, the AI saw drugs like Propofol and Rocuronium and cheerfully hallucinated standard wellness advice like "take with food." For intravenous general anesthetics and paralytics administered during surgery, that advice is dangerously wrong.

To solve this, I brought in external adversarial AI auditors (Sol and DeepSeek) to relentlessly attack our codebase. Through their critique, we engineered a dedicated classification step in the pipeline that strictly separates in-hospital procedural medications from post-discharge care.

Latency was another hurdle. Early experiments with 550B models took 30 to 40 seconds per bill. By switching to NVIDIA's Nemotron 30B MoE model on Nebius Token Factory, we slashed end-to-end reasoning latency down to 8.2 seconds, making the app genuinely usable.

Accomplishments that I'm proud of

• As a non-developer, orchestrating multiple AI agents to deliver a fully functional, verifiable healthcare prototype from scratch.
• Successfully processing authentic, dense US surgical bills with 15+ clinical items and achieving accurate clinical separation in under 9 seconds.
• Zero English UI leakage across all user stages, proven through automated browser recording and adversarial cross-validation.

What I learned

Directing AI agents taught me that prompt engineering alone is never enough for healthcare. AI models must be anchored to deterministic parsers and subjected to strict adversarial critique. NVIDIA's Nemotron 30B MoE model proved remarkably efficient at structured clinical reasoning when given proper architectural guardrails.

What's next for PainFree

PainFree is architected to scale from a prototype into a full patient companion mobile app:

1. Automated Health Sync: Direct background integration with Apple HealthKit and Android Health Connect to automatically track daily activity drops and sleep disturbances without requiring manual entry.
2. Mobile Camera Document Scanner: Instant camera scanning at hospital billing windows with on-device perspective correction and glare removal before uploading to Nebius Token Factory for real-time ~8s bill audits.
3. Doctor 1-Page Handout & QR Share: Export clean clinical 1-page reports or generate local offline QR codes for instant sharing during 3-minute outpatient doctor consultations.
4. Multi-regional Billing Support: Expand deterministic parser rules to support UK NHS prescription statements, Korean Health Insurance Review & Assessment (HIRA) itemized non-reimbursable statements, and German GKV billing formats.
5. Zero-Knowledge Patient Vault: End-to-end client-side encryption for all stored pain telemetry and scanned medical bills to guarantee patient privacy and strict HIPAA/GDPR compliance.

Built With

• nebius
• nvidia
• token