Sanket

Complete system on GitHub 👇 github.com/sanketughadmat… Building in public works. 🚀 Day 30 ✅

What would I do differently? 1. Add evaluation on Day 1, not Day 21 2. Docker from the start 3. Test with real data, not toy examples 4. Track cost from Day 1 5. Load test earlier Learn from my mistakes.

Day 30 of building Agentic AI systems. 30 days. 30 commits. Production-ready. From "I should learn RAG" to deployed Docker container with multi-provider failover, evaluation, and self-correction. Final thread 🧵 #AgenticAI #RAG #BuildInPublic

Built Self-RAG with iterative improvement 👇 github.com/sanketughadmat… Iteration 1: Score 0.72 Iteration 2: Score 0.88 ✨ +22% improvement

Tested Regular RAG vs Self-RAG 👇 Regular RAG: 0.70 Self-RAG: 0.80 +14.3% quality improvement. Self-RAG adds reflection + iterative retrieval. More tokens. More latency. But when correctness matters > speed, it’s absolutely worth it.

Day 29 of building Agentic AI systems. The first answer is a guess. The second answer is informed. The third answer is refined. Today: Self-RAG - systems that critique and improve their own responses. #AgenticAI #RAG #BuildInPublic

Built production-ready RAG API 👇 github.com/sanketughadmat… docker-compose up -d curl localhost:8000/health open localhost:8000/docs From prototype to production in 28 days.

Built a complete deployment stack: - FastAPI REST API (async) - Docker container (reproducible) - Health checks (automatic monitoring) - Metrics endpoint (observability) - CORS middleware (cross-origin) Everything in one container.

Day 28 of building Agentic AI systems. "If it's not containerized, it's not production-ready." Today: Deployment - FastAPI + Docker + docker-compose. #AgenticAI #RAG #BuildInPublic

Built multi-provider RAG with intelligent failover 👇 github.com/sanketughadmat… Primary fails → auto-switch to secondary Rate limit hit → failover (not retry) All providers down → graceful degradation Thanks @thedarshanjoshi @AustenMakers for the insights 🙏

The cooldown pattern: Provider fails 2x in 10s → bench for 5 mins Why? - Prevents retry storms - Saves API costs on known-bad providers - Gives service time to recover h/t @AustenMakers for this one.

Day 27 of building Agentic AI systems. A single point of failure is a single point of failure. Today: Multi-provider resilience - inspired by real production insights from @thedarshanjoshi and @AustenMakers #AgenticAI #RAG #BuildInPublic