CS student at Texas A&M in Austin, TX — B.S. December 2027 (3.87 GPA, math minor). I build and ship full-stack products end to end, work on application security for a production multi-tenant SaaS platform, and spent a summer writing an LLVM compiler backend for a vector DSP.
Live: brianfu.vercel.app · Axle
$ security --summary
50+ production vulnerabilities remediated access-control, injection, race conditions
13-stage pre-push verification pipeline negative RLS probes, regression suites
cross-tenant authorization gap closed proven by reverse-reproduction
| project | what it is | |
|---|---|---|
▸ |
Axle — live ↗ | AI-powered used-car search — plain English in, ranked real listings out. LLM pipeline with schema sanitization, fuzzy alias matching, two-tier model routing. Deployed serverless: FastAPI on Lambda behind API Gateway, React on CloudFront/S3, DynamoDB cache with native TTL holding it inside a 1,000-call/month budget.React FastAPI AWS Lambda DynamoDB API Gateway PostgreSQL |
▸ |
personal-website — live ↗ | Generative portfolio — a hand-written WebGL fragment shader (domain-warped fBm) that warps toward your cursor, with spring-driven scroll-linked reveals.TypeScript React WebGL GLSL |
▸ |
project-horizon | Narrative exploration game, solo since 2023 — dialogue engine with typewriter text synced to voiceover audio and branching choices that gate progression.C# Unity |
▸ |
CarStatus | CLI that pulls a BMW's live odometer, fuel/battery, and location via the Smartcar API.Node.js REST |
▸ |
wild-west-party-game | Live-multiplayer party game built in 24h at HowdyHack 2024 — I built the frontend.TypeScript Next.js |
How Axle turns a plain-English question into ranked listings — and stays inside a 1,000-call/month model budget.
flowchart LR
Q["Plain-English<br/>query"] --> CDN["React<br/>CloudFront / S3"]
CDN --> GW["API Gateway"]
GW --> API["FastAPI<br/>on Lambda"]
API --> CACHE{"Multi-tier cache<br/>memory / disk / DynamoDB TTL"}
CACHE -->|hit| RANK["Ranked listings"]
CACHE -->|miss| SAN["Schema<br/>sanitization"]
SAN --> ALIAS["40-entry alias map<br/>+ fuzzy matching"]
ALIAS --> ROUTE{"Two-tier<br/>model routing"}
ROUTE -->|routine parse| SMALL["Cheap model"]
ROUTE -->|ambiguous| LARGE["Capable model"]
SMALL --> RELAX["10-level constraint<br/>relaxation + scoring"]
LARGE --> RELAX
RELAX --> RANK
RANK --> CDN
Two things hold the budget: repeat questions never reach a model, and only genuinely ambiguous ones reach the expensive one. The relaxation ladder is why a query with no exact match still returns ranked results instead of an empty page.
 |
 |
 |
