Retrieval architecture that indexes generated questions instead of raw document chunks
Cosine similarity 0.71 → 0.94 · 75%+ token reduction · Independently validated by a Seoul National University research team — arXiv:2508.09755
PythonQdrantOpenAIRAG
Creator platform — built solo, end-to-end
19k+ MAU · 1.2M+ monthly DB queries at sub-150ms latency · ~$38/month infra · Razorpay payments integration
Next.jsPostgreSQLNode.jsTypeScriptRailway
Multi-LLM ETL pipeline for YouTube content transformation
GPT, Claude, and Gemini coordinated via pg-boss queues across a 5-node Railway deployment. Schema v2 preserves 6 fields across regeneration calls to prevent hallucination from partial context.
TypeScriptpg-bossRailwayOpenAIClaudeGemini
AI support widget built on QIndex-RAG
Confidence-gated responses suppress outputs below 0.70 cosine similarity. In progress: streaming · citations · RBAC · prompt versioning · eval harness · observability.
TypeScriptNext.jsQdrantPostgreSQLExpress
| Regify | CLI registry generator via ts-morph AST analysis · 100+ npm downloads |
| Fysk | shadcn-CLI-compatible React component library |
| STL | Framework-agnostic DSL for advanced table composition |
| Languages | TypeScript · Python · SQL · JavaScript |
| Backend | Node.js · Express · PostgreSQL · pg-boss |
| Frontend | Next.js App Router · React · Tailwind CSS |
| AI / ML | RAG pipelines · Qdrant · OpenAI · Anthropic APIs · Embeddings · Eval design |
| Infrastructure | Railway · Vercel · Docker · GitHub Actions · AWS S3 |