AI Express proposes one credit balance for multiple AI tools

Instead of maintaining separate subscriptions for different AI tasks, could users draw from one shared credit balance? AI Express explores that approach through a unified interface and independently deployable services. Its reported results are encouraging, but the preprint leaves important questions about evaluation and cost comparisons unresolved.

AI Express brings several AI services into a single interface with a shared, usage-based credit system. In an arXiv preprint, Deepit Sapru of the University of Illinois Urbana-Champaign describes the platform’s architecture and reports performance, usability and billing tests. The central proposition is integration: combining tools and payments rather than introducing a new general-purpose AI model.

Sapru frames the project as a response to fragmented services, recurring subscription costs and the disruption of switching between interfaces. AI Express is intended to make varied capabilities accessible to students, independent developers and other users without requiring specialist technical knowledge. Those accessibility and affordability benefits are goals supported by some reported observations, not established outcomes for every user.

How the services fit together

The described frontend uses React and TypeScript, while Node.js coordinates backend requests. An API gateway authenticates users, validates requests and applies rate limits. NGINX handles encrypted connection termination and load balancing. PostgreSQL stores account information, service pricing, content and credit transactions; Redis supports session caching and rate limiting.

Separate microservices handle language and code, image processing, and structured data analysis. This separation is intended to let services be deployed and scaled independently. The paper describes a mixture of locally hosted and externally accessed AI engines, but does not identify specific model versions or the hardware used for its performance measurements.

The code reviewer retrieves a GitHub repository and uses language-model analysis to flag problems and suggest changes. The data analyst accepts CSV, TSV and Excel files, translating natural-language requests into SQL or Python scripts. Writing tools generate scripts or documentation and refine grammar, word choice and tone. An image service uses inpainting to reconstruct regions obscured by visible watermarks. Resume building and web scraping are also described. A text-to-website generator appears in the abstract, but receives no comparable implementation or evaluation detail.

Users purchase credits, and service charges are tied to computational weights. Sapru reports concurrent testing of the PostgreSQL credit engine and says deductions matched computational effort. However, the paper does not provide a detailed service tariff or enough transaction-test results to independently assess billing reliability across failure cases.

What the evaluation reports

Average end-to-end latency ranged from 600 milliseconds for grammar checking with a 5 KB text payload to 3,100 milliseconds for watermark removal with a 5 MB image. Script generation averaged 850 milliseconds, code review 1,250 milliseconds and data analysis 2,400 milliseconds, each with a different average payload size. These are reported measurements for those workloads, not guarantees for arbitrary inputs.

Using Apache JMeter, the author reports handling 500 concurrent users for 10 minutes with 99.2% availability and an 18% increase in average response time at peak load. A separate stress test reportedly sustained 1,000 concurrent code-review requests while maintaining stability. The paper describes an autoscaling threshold of 70% CPU utilization, but gives limited infrastructure and workload detail for reproducing these results.

Output-quality findings are less specific. The reviewer reportedly identified syntax problems and refactoring opportunities in sample Python and JavaScript repositories. The analytics agent is described as having a high success rate when tables were well formatted. Image reconstruction reportedly achieved high structural similarity, with minor blurring in complex textures. Numerical success rates, similarity scores and detailed benchmark datasets are not supplied, limiting assessment of these claims.

Usability and savings need context

Fifty student beta users gave the platform an overall rating of 4.2 out of 5. The paper also reports an onboarding rating of 4.3 out of 5, a 94% task-completion rate and a System Usability Scale score of 82. These results describe that student evaluation, not a demonstrated experience across broader user populations.

A representative student workflow involving code review, data analysis and resume optimization consumed 340 credits, priced at $3.40. Sapru compares this with $45 in monthly subscriptions and reports a 92% saving. That example compares selected usage charges with fixed subscription costs; it does not establish equivalent savings at other usage levels. The feature-and-price comparison likewise is not a controlled test of competing tools’ output quality.

AI Express therefore offers a concrete account of integrating AI services and shared billing, with preliminary reported operational results. Collaboration, client-side models, audio tools, mobile applications and other extensions remain future proposals. As a preprint, the work should be read as an author-reported system evaluation rather than independently validated evidence of superior quality or universally lower costs.