ClaimIQ

Inspiration

Medical billing errors cost billions annually, often driven by trivial administrative mismatches. For instance, a hospital using a DD/MM date format on a discharge summary alongside an MM/DD format on a lab report easily triggers automated or manual "timeline discrepancy" flags, leading to immediate insurance denials for valid claims. We realized that healthcare compliance shouldn't rely on overworked human reviewers catching cross-document data collisions. We built ClaimIQ to ruthlessly audit multi-document packages, normalizing data formats and catching clerical errors before a hospital ever presses 'submit'.

What it does

ClaimIQ is an automated healthcare insurance auditing web application where users upload multi-document claim packages for pre-submission intelligence. Think of it as Grammarly for insurance claims.

The platform digitizes and classifies uploaded documents, then leverages AI to cross-verify patient identity, medical reasoning, and billing consistency across files. Instead of throwing a generic error, it generates a weighted Claim Readiness Report with prioritized actionable fixes. For example, the AI explicitly natively catches contradictions, such as flagging when anti-diabetic medications (Insulin/Metformin) are billed on a prescription, but the lab reports indicate a normal fasting blood sugar of 96 mg/dL.

How we built it

We engineered a robust, real-time architecture utilizing a 5-stage intelligent processing pipeline.

Frontend: We built the client using React 19 and Vite, integrated Lucide React for UI icons, and styled it with custom modular Vanilla CSS.
Backend & Streaming: The core logic runs on Python 3 via FastAPI and Uvicorn. To handle long-running document processing without blocking the UI, we implemented Server-Sent Events (SSE) via sse-starlette to stream real-time audit updates.
Document OCR: We utilized Docling for primary document-to-markdown conversion, paired with an automated fallback to pypdf for messy scans.
Medical AI Node: We integrated the DeepSeek API (deepseek-chat model) via the OpenAI Python SDK. We run the model at a low temperature of 0.1 with json_object enforcement to ensure strict, reproducible clinical logic without hallucinations.

To evaluate claim safety mathematically, we engineered a 5-pillar weighted scoring algorithm. The overall readiness score is calculated using the following formula:

$$S=\sum_{i=1}^{5}(P_i\times W_i)-\text{Penalties}$$

Where the pillars $P_i$ and their respective weights $W_i$ belong to the set $W_i \in {0.25, 0.20, 0.20, 0.20, 0.15}$ mapped to Clinical & Medical Logic, Document Completeness, Identity Verification, Financial & Billing Accuracy, and Timeline & Date Integrity.

We also built a Hybrid Cross-Document Verification Engine. To catch subtle patient name mismatches without triggering false rejections, we merge Python rules with string similarity algorithms, enforcing:

$$\text{FUZZY_THRESHOLD}=0.85$$

Additionally, we apply a 10.0 rupee tolerance when verifying net hospital bill totals against insurance claim forms to bypass harmless rounding differences.

Challenges we ran into

Processing massive, multi-page medical documents initially caused severe LLM latency and exorbitant API token costs. We quickly realized that sending a 10,000-character raw OCR dump was unscalable.

To minimize latency and token overhead, we engineered a brutal pre-trimming mechanism for our input payloads. We capped quick classification snippets at 1,500 characters and restricted deep clinical extraction texts to just 3,500 characters, exclusively filtering for complex document types.

Accomplishments that we're proud of

Lightning Fast Execution: The entire 5-page multi-document pipeline executes end-to-end in under 30 seconds.
Extreme Cost Efficiency: By slashing token usage by ~15,500 tokens per claim, the total AI processing cost per multi-document claim package is driven down to just $0.003 (approx. ₹0.25), making it highly profitable to run at scale.
Dynamic Financial Risk Engine: A critical document failure immediately deducts 35 points, dynamically placing exactly 35% of the claim amount at risk. For example, a missing discharge summary on an ₹85,000 claim mathematically isolates and flags ₹29,750 as active financial exposure.

What we learned

We learned that insurance claims aren’t primarily rejected because of actual medical fraud; they are rejected because human reviewers are forced to manually cross-reference dense PDFs hunting for date mismatches and missing signatures. Traditional software with rigid IF/THEN logic cannot fix this; solving it requires true contextual medical reasoning.

What's next for ClaimIQ

Pre-Authorization Checker: Shifting the audit even further left to pre-approve procedures before they happen, saving hundreds of administrative hours.
Fraud & Overcharge Detection (TAM): Expanding our deterministic checks against historical data to spot duplicate claims and CPT overbilling.
Automated Appeal Letter Generator: Using the LLM to draft evidence-backed appeal letters for claims that are unfairly rejected by insurers.

Track 01 Money, Jobs & AI — Will AI Steal My Future? [Audit Bot]