| 99.8% | 24 | 86% | 1,000+ |
|---|---|---|---|
| Uptime sustained for 2,500+ businesses | Vulnerability patterns detected by VibeSec | Hallucination detection accuracy | Engineering readers reached |
I'm a Software Engineer at Udeck Services, architecting backend systems and cloud infrastructure that sustain 99.8% uptime for 2,500+ businesses managing 10,000+ SKUs — cutting p99 latency by 40% along the way.
I work at the intersection of backend engineering and applied AI: building systems that hold up under real load, not just demos. That curiosity turned into shipped tools — a scanner that catches what AI-generated code gets wrong, a framework that judges whether LLMs can be trusted, and pipelines that fail gracefully instead of silently.
const ayush = {
role: "Software Engineer @ Udeck Services",
stack: ["Python", "Node.js", "PostgreSQL", "Redis", "AWS"],
building: "production-grade reliability into AI-backed systems",
shipped: ["VibeSec (PyPI)", "LLM Eval Framework", "Data Extractor"],
alwaysLearning: true,
};
🛡️ VibeSecSecurity scanner for AI-generated code Catches what Cursor, Bolt, and Claude Code miss — hardcoded secrets, disabled RLS, hallucinated packages, XSS — via deterministic AST analysis + LLM-powered fixes.
|
Three-tier judge architecture for LLM trust A distributed judge chain — models evaluating models — benchmarking factual accuracy, safety, and adversarial robustness.
|
|
AI business intelligence pipeline at scale Scrapes 8 public data streams through an async LLM fallback chain (Qwen-2.5-VL → Llama-3.1 → Regex) for extraction that degrades gracefully, never silently. |
Fedora · GirlScript Summer of Code · Medium Shipping production code and architecture docs to Fedora's DRI repos, and writing technical essays on ML systems and full-stack engineering. |
