Building an On-Premise Health Insurance Appeals Generation System

healthinsurancellm · Healthcare · 2023 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Health insurance appeal generation

The project aimed to help patients draft appeal letters for denied health-insurance claims using a low-cost, self-hosted LLM system, while acknowledging that the prototype was not production-ready and operated in a legally sensitive domain.

Constraints:

• Health-insurance appeal text is high-stakes and jurisdiction-specific.
• The system had to be built on a personal budget using available public records and consumer hardware.
• The model needed domain adaptation from independent medical review records.
• The speaker explicitly framed the work as early and not production-ready.

Architecture

The project uses California independent medical review board records to create synthetic fine-tuning data, fine-tunes a language model cheaply on rented Lambda Labs compute, and serves inference on a self-hosted Kubernetes cluster using consumer GPU hardware and vLLM.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  patient-claim-context[Patient Claim Context] --> appeal-generator[Appeal Generator]
  california-review-records[California Review Records] --> synthetic-data-pipeline[Synthetic Data Pipeline]
  synthetic-data-pipeline --> fine-tuned-model[Fine-Tuned Model]
  fine-tuned-model --> vllm-server[vLLM Server]
  vllm-server --> self-hosted-kubernetes[Self-Hosted Kubernetes]
  self-hosted-kubernetes --> rtx-3090-gpu[RTX 3090 GPU]
  appeal-generator --> fine-tuned-model[Fine-Tuned Model]

Components

Data flow

Public California review-board records are converted into synthetic training examples, used for fine-tuning, and then served through a self-hosted Kubernetes/vLLM stack on consumer GPU hardware to generate appeal-letter drafts.

Why this architecture

The architecture was driven by budget and control: public review records provided domain data, synthetic examples made fine-tuning possible, Lambda Labs handled cheap training, and self-hosted Kubernetes/vLLM kept inference on personal hardware.

Models used: fine-tuned language model · vLLM

Scale

• The cited insurer denied around 20% of claims in 2019.
  “According to the presentation, Anthem Blue Cross denied approximately 20% of claims in 2019, and there are allegations that some insurers use AI to automate claim denials.”summary, stated
• Fine-tuning cost about $12 using Lambda Labs.
  “While still in early development and explicitly not production-ready, the project demonstrates a low-budget approach to LLMOps with fine-tuning costs around $12 using Lambda Labs and inference running on personal hardware costing approximately $748.”summary, stated
• Personal inference hardware cost approximately $748.
  “While still in early development and explicitly not production-ready, the project demonstrates a low-budget approach to LLMOps with fine-tuning costs around $12 using Lambda Labs and inference running on personal hardware costing approximately $748.”summary, stated
• The deployment used an RTX 3090 GPU.
  “The solution involves fine-tuning a language model on synthetic data derived from California's independent medical review board records, deploying it on a self-hosted Kubernetes cluster with consumer GPU hardware (RTX 3090), and serving it via vLLM.”summary, stated

Alternatives considered (roads not taken)

Use a hosted production SaaS deployment.

Rejected because: The project was a personal, low-budget prototype using self-hosted hardware.

Rely on raw general-purpose LLM behavior.

Rejected because: The appeal domain needed fine-tuning on review-board-derived examples.

Treat the prototype as production-ready.

Rejected because: The source explicitly says it is early and not ready for production use.

Patterns

synthetic-data-generation (training)

Public review records are converted into synthetic examples for fine-tuning.

fine-tuning-adaptation (domain adaptation)

The model is fine-tuned for the health-insurance appeal domain.

privacy-preserving-processing (deployment)

The prototype favors on-premise/self-hosted inference rather than external SaaS processing.

cost-aware-routing (infrastructure)

The system was designed around low fine-tuning and hardware costs.

Tradeoffs

Training data

Chose Synthetic data from California review records over Manual collection of appeal-letter training data.

Rationale: The public records were available and structured enough to bootstrap fine-tuning.

Serving environment

Chose Self-hosted Kubernetes and vLLM on RTX 3090 over Managed cloud inference.

Rationale: The project emphasized low-budget personal control of inference.

Project maturity

Chose Prototype disclosure over Production legal/medical assistant positioning.

Rationale: The model quality and domain risk do not justify production claims.

Gotchas (what bit them)

Low-cost domain adaptation can still produce only prototype behavior.

Root cause: Synthetic data and small-scale fine-tuning may not capture legal and medical complexity.

Resolution: Keep the system framed as early development rather than production-ready.

Consumer GPU deployment has hardware constraints.

Root cause: The inference plan depends on specific local GPU hardware.

Resolution: Design the deployment around the available RTX 3090 and self-hosted Kubernetes environment.

Public records can create jurisdictional bias.

Root cause: The current data source is California-specific.

Resolution: Treat the California review records as an initial data source, not complete national coverage.

Evaluation

The source reports qualitative prototype evaluation rather than a formal benchmark: the fine-tuned model works okay but not amazingly, with cost, hardware, and data constraints emphasized.

Metrics: Fine-tuning cost · Hardware cost · Prototype quality · Claim denial context · Synthetic data feasibility

Outcomes

• The project produced an early appeal-letter generation prototype.
• It demonstrated fine-tuning on synthetic review-board-derived data.
• It showed that low-budget self-hosted LLMOps is possible but not production-ready.

The summary reports around $12 fine-tuning cost, approximately $748 personal hardware, RTX 3090 inference, and the motivating 20% claim denial statistic.