CrisisBridge

What We Built

USignal is an artificial intelligence-based crisis prevention system that is able to detect early. Predictive indicators of financial disaster, prior to the financial catastrophe. USignal is basically one trigger event, such as an unanticipated medical office visit, to one message, such as a text alert to your phone. Models the domino effect of effects through a household, and timeline. It provides three outputs: a Collapse Risk Score. A Crisis Forecast Timeline and a personalized Action Plan. We also developed PayPlan which is a medical bill analyzer which checks CPT. Procedure codes are against the real medical databases and flags suspicious charges. Identifies and provides a clear explanation of like duplicates, inflated units and price outliers. VERDICT: PAY, PAUSE, or DISPUTE. From there, PayPlan AI Creates a professional dispute letter that the user can e-mail to the hospital immediately.

What Inspired Us

We continued to ask ourselves the same question when brainstorming: why do When the warning signs are there, people find themselves in their full financial crisis. weeks earlier? Having a $1,500 medical bill does not make somebody instantly homeless. But it starts a chain missed rent, skipping meals, not collecting medicine, notices and by the time someone calls for assistance the damage is already done. All of the community support systems we observed were designed to REACT. Nobody had constructed something to foretell. This was the gap that was filled by USignal.

The medical bill came as a surprise. The real bills were reviewed, and realized that most people have no idea what a CPT code is, whether they've ever, They have had their bills doubled for the same service, or they have a legal right to complain about it.They are billed over for the same service twice or they have a legal right to complain. To contest an accusation. It seemed a problem that could be solved, so we did.

How We Built It

It's a single page React app that uses the Anthropic Claude API 4-6-sonnet. The AI layer We use Claude in four ways: Crisis classification reading a plain-language description and Identifying those needs that are vulnerable (housing, food, legal, medical)

• Risk scoring and cascade modeling risk estimates of collapse probability Creating and predicting stress distributions over time context and flagging errors in CPT coding, and interpreting the context of CPT coding. generation of dispute letters, diagnosis to visit mismatches
• PayPlan AI letters Generating legitimate, targeted dispute and Bill data, with appeal letters. The code validation layer For MediClear, we're using two APIs that are free and no-key from the The National Library of Medicine (NLM):
• HCPCS/CPT code lookup:
  https://clinicaltables.nlm.nih.gov/api/hcpcs/v3/search Use the lookup feature to find diagnosis codes in ICD-10-CM.
  https://clinicaltables.nlm.nih.gov/api/icd10cm/v3/search These are used to validate all the codes on the bill as they type — invalid codes and filling description, results, Before Claude sees anything, these are put into the rule-based checks. The rule-based checks We do some deterministic checks on the client side prior to the AI call: | Check | Trigger | Invalid code | NLM does not return any answer | Duplicate charge: This is when a charge appears twice with the same CPT, same date, and same amount. | Units in excess of | > Maximum billing units | If the price outlier is the biller: Billed amount is greater than 3 times the benchmark | EOB mismatch | Provider bill ≠ insurance EOB | Surprise visit | Emergency visit with high patient responsibility | These generate a Bill Risk Score of 0–100: The Risk Score is calculated by summing the flag weights and capped at 100. Score bands:
• 0–25 → Low risk
• 26–60 → Needs review
  
• 61–100 → High risk — dispute recommended

The forecast timeline

The Crisis Forecast Timeline is created using Claude and the information provided to it. The factors include user's income, bill amount, existing obligations and household size. It provides a mapping of day to day consequences so users can see what's coming and Act at the scene before it gets here.

Challenges We Faced

Making the AI specific not general.
Although the original Claude prompt yielded suggestions such as 'consider', the initial iterations were a bit incoherent. contacting a financial advisor." This is not going to help a person in crisis. We Repeatedly edited the system prompt so that all the output was specific real organization names, real steps, real deadlines. CPT code coverage. The NLM HCPCS API supports the majority of codes but not all private payer codes. We've been good, if a code is not found, we mark it as such. Say something about it and give a reason for why.

Making the risk score trustworthy. Numbers such as an 87% collapse risk are very unsettling or vague. Depending on the presentation, it will create a dependence. We devoted a lot of time to the explanation layer: each score gives a detailed explanation of factors that contributed to the score and what you can do about each of these. Simplifying for those who need help. Our users aren't technoliterate, they are terrified, lost, and confused. All design choices were made with the intent of decreasing friction, not increasing it. No login. No 20-question form. This is a form with one text box, two follow-up fields and a complete plan.

What We Learned

Medical billing is truly broken the way most people don't realize. Until anyone suffers from a mischarged cost that they can't afford The difficulty of AI product design isn't the model, it's getting to know -- Avoid asking questions of the user

• A crisis should be prevented, which involves emotional design (calm, clear, and responsible), as well as technical design (appropriate, well organized, and feasible). Practical experience, and strict validation and real data (trustworthy) and technical depth (real) cannot replace one with the other (letters). The APIs provided by the NLM are free and open to the public, and are absolutely amazing. Underused in Civic Tech

What's Next

Live resource database – up to date capacity data from partner organizations For users who don't have a reliable internet connection: orgs and people are located.- Org dashboard, anonymized community trends, where The risk of financial collapse is greatest in each ZIP code

• Expanded code coverage, NCCI edits, modifier validation, When API access is available, and CMS pricing benchmarks available.

Built With

• React
• Tailwind CSS The Claude-Sonnet-4-6 model provides the following features:The Claude-Sonnet-4-6 model supports the following features:
• NLM Clinical Tables API (HCPCS + ICD-10-CM)
• JavaScript

Built With

• chatgpt
• claude