TexFolio's AI LaTeX resume builder compiles PDFs with pdflatex

TexFolio released an open-source, AI-powered LaTeX resume builder that compiles PDFs with pdflatex and scores each resume on Content, ATS compatibility, Format, and Impact via a LangGraph multi-agent pipeline [DevTo]. The product ships as a Turborepo monorepo with three apps: a React 19 front-end, a Hono v4 API, and a dedicated Docker container for pdflatex [GitHub].

The API enforces a strict middleware chain—request ID, structured JSON logging, CSP headers, CORS, and a tiered Redis-backed rate limiter (300 req/min for paid plans, 60 req/min for free, 20 req/min for anonymous users). Authentication is handled by Clerk JWTs, and role-based access control (RBAC) is enforced via a requireRole middleware that compares numeric role weights (Owner 4 → Viewer 1).

Heavy LaTeX compilation is offloaded to BullMQ workers. The queue runs with concurrency: 2 and a limiter of 5 jobs per minute, each job retrying up to three times with exponential back-off (2 s → 4 s → 8 s). Workers spawn a Docker container and invoke pdflatex via spawn with an argument array, eliminating shell injection risk [DevTo].

The AI scoring pipeline uses LangGraph nodes—Content (30 % weight), ATS (25 %), Format (20 %), Impact (25 %)—each backed by a prioritized LLM provider chain: NVIDIA NIM (Llama 3.1 70B) → Google Gemini 1.5 Flash → Groq (Llama 3.1 70B) [GitHub].

TexFolio's architecture provides a reproducible blueprint for web services that must run untrusted code, such as code-execution sandboxes and document converters. The combination of BullMQ queueing, Docker isolation, and explicit spawn-based command execution ensures secure and performant LaTeX rendering at scale.

The layered provider fallback and per-node “continue on failure” strategy turns inherently flaky LLM calls into a dependable feature, demonstrating that production-grade AI can be achieved with classic resilience patterns [DevTo].