WEBSITE by Aras
Back

Aerrus Use Case

aras.aniulis.dev

A personal portfolio and public AI assistant that turns Aras's project notes, career history, and engineering approach into a polished, grounded web experience.

What it is

A portfolio with a public, cost-aware AI assistant.

aras.aniulis.dev is the personal identity and case-study surface in the Aerrus ecosystem: a multi-section portfolio for projects, career history, education, and how Aras works, backed by the same engineering discipline as a product app.

The AI assistant is publicly available. Visitors do not need to sign in to ask questions; the chat is exposed as part of the portfolio and answers from Aras's own document corpus instead of open-ended model memory.

Under the hood, GitLab CI ingests Google Drive documents, extracts text, splits it into overlapping chunks, and stores both Azure OpenAI embeddings and full-text searchable context in Azure Cosmos DB. The corpus is indexed for semantic vector search and literal keyword search, so short prompts such as a single topic can still find the right project or thesis material.

At request time, the Vercel-hosted website keeps a read-only Cosmos credential, asks the chat model for compact retrieval terms, embeds the expanded semantic query, and retrieves context with Cosmos DB hybrid ranking across vector distance and full-text score. The answer model only receives the selected excerpts and responds in a friendly, grounded voice.

The public chat is deliberately rate-limited before any Azure OpenAI call: 10 requests per minute per IP, 20 requests per minute globally, 100 requests per UTC day globally, and 1000 requests per UTC month globally. Those caps exist because every public message can trigger query expansion, paid embedding, hybrid retrieval, and chat-generation work; without hard limits, automated traffic or a simple popularity spike could create disproportionate cloud cost.

Aerrus Platform SSO is still wired into the portfolio for identity, session display, logout, and future protected surfaces, but the current AI chat and public portfolio pages are open to anyone. Cost control and abuse resistance come from Redis-backed throttling rather than a sign-in gate.

Hybrid grounded chat

The assistant is available without login and retrieves from Aras's Drive-backed corpus with query expansion, vector search, and full-text search.

Cost-aware rate limits

Redis enforces per-IP, global minute, global day, and global month limits before paid Azure OpenAI requests are made.

Clean trust boundary

Ingestion credentials live in GitLab CI, while the public app only holds read access for retrieval and cannot mutate the knowledge base.

Live Site

Open the deployed portfolio, try the public AI chat, or compare it with the underlying Aerrus platform case study.