Case Study
ebia-insights
A production analytics and semantic-search platform — RAG over a real corpus, statistically grounded trends, and data that refreshes without a redeploy.
EB-1A immigration appeals intelligence · Live · 2026
By the numbers
The shape of the system
Overview
What it is, and the problem it solves
A gated intelligence dashboard over thousands of federal immigration appeal decisions. It pairs natural-language semantic search with criterion-level trend analytics, built as a clean, read-only consumer of an upstream data pipeline.
Practitioners reasoning about EB-1A petitions have no good way to ask, in plain language, what has actually persuaded the appeals office — or to see which arguments are gaining or losing ground over time. The raw decisions are unstructured, voluminous, and refresh continuously. The platform turns that corpus into searchable, trend-aware intelligence that stays current without manual intervention.
Engineering
Technical highlights
The decisions and techniques that make the system fast, current, and trustworthy.
Production RAG over a real corpus
Natural-language queries are embedded with Gemini (768-dim, retrieval-tuned), matched against a Vertex AI Vector Search index, then re-ranked with a hybrid relevance-plus-recency score (exponential decay, ~2-year half-life, normalized so semantic relevance still dominates). A lexical fallback keeps search working even when the embedding or vector service is unavailable.
Zero-redeploy data architecture
Every request checks a tiny version manifest; when the upstream pipeline publishes a new snapshot, the API loads the JSONL artifacts from object storage in parallel into in-memory maps and hot-swaps them. Fresh data goes live in seconds with no deploy, no cache-bust, and no downtime.
Shipped full-stack, solo
A React 19 + Vite SPA, an Express API on Cloud Run, Firebase authentication, and a Firestore beta-access gate enforced with atomic transactions — plus a custom design system with light/dark themes and reduced-motion support. One person, the whole stack, in production.
Analytics you can defend
Criterion trends are classified with a two-proportion z-test on volume-normalized shares (not raw counts), compared over fixed 12-month windows to eliminate partial-year bias, and gated on significance, sample size, and effect size — producing rising / falling / steady labels that hold up to scrutiny.
Architecture
How a request flows
Stack
Tech stack
Frontend
Backend
AI / Retrieval
Cloud / Infra
Quality
Why this matters for your team
What this project demonstrates
Applied AI that survives contact with real data
I build retrieval systems — embeddings, vector search, hybrid ranking, sensible fallbacks — that work on messy production data, not curated notebook demos.
Architecture designed for operations
Data that refreshes without redeploys, least-privilege service accounts, atomic access control, and memory and cost sized to the workload. I design for the day after launch.
End-to-end delivery
SPA, API, authentication, infrastructure, tests, and documentation — shipped by one engineer. A single engagement can move the whole system, not just one layer.
Next step
Building something similar?
Bring the problem, the constraints, and the timeline. I’ll help you design and ship it.