Prashant Kafle

Computer Engineering| Backend Developer (DRF)|Web Scrapping

[Prashant Kafle @PrashantK7738]

Try it live: https://t.co/cFOjlIc7ML

Code: https://t.co/UDzHz4NmXa

Would love feedback from anyone working in RAG/LLM systems.

#AIEngineering #RAG #BuildInPublic

[Prashant Kafle @PrashantK7738]

@Areyaarkya @mr_h8wks what do you think?

[Prashant Kafle @PrashantK7738]

Stack: React + Vite, FastAPI, LangGraph, PostgreSQL, Qwen 3 27B for inference, with an Ollama fine-tuned model path for fully offline operation.

🔗 Try it: https://t.co/oaeI4HWypb

💻 Code: https://t.co/87gDKQ5Jvb

#LangGraph #MultiAgent #AIEngineering

[Prashant Kafle @PrashantK7738]

Built a 7-agent LangGraph pipeline that turns messy financial/sales CSVs into a full analytical report — with only 2 of the 7 agents actually calling an LLM.

Here's why that ratio matters 🧵

[Prashant Kafle @PrashantK7738]

Output Validation sits between Insight Generation and Report Generation specifically to scope LLM retries — if the insights don't check out, I retry just that step, not the whole pipeline.

[Prashant Kafle @PrashantK7738]

Full paper + code (ROUGE-L, Bradley-Terry, SelfCheckGPT, kappa — all from scratch):
🔗 Paper: https://t.co/hzLjXbt1CT
🔗 Code: https://t.co/GAji97Tmuh
#AIAgents #LLMEval #MachineLearning

[Prashant Kafle @PrashantK7738]

Everyone's shipping AI agents. Almost no one is checking if their outputs are actually correct.

I spent a week building an eval pipeline to test this — and found something that should worry anyone using "LLM-as-a-judge."

[Prashant Kafle @PrashantK7738]

Lesson:
reliability (a consistent score) ≠ validity (the correct label).
If you're using an LLM judge in your pipeline, check per-class not just one aggregate kappa.

[Prashant Kafle @PrashantK7738]

@MessiXtraHQ The North Forgave you😇

[Prashant Kafle @PrashantK7738]

@paud73496 @ReshadRahman @JijantesFC ��

[Prashant Kafle @PrashantK7738]

@FabrizioRomano @StreamviDL download this please

[Prashant Kafle @PrashantK7738]

@jaydenGTM9009 Thanks! Always open to good ideas, feel free to drop them here

[Prashant Kafle @PrashantK7738]

I stopped reading PDFs. I started talking to them instead.
Built Contextra AI — a RAG system that turns any document into a knowledge base you can chat with, summarize, and turn into flashcards.
Full-stack, live demo, open source 🧵

[Chess.com @chesscom]

never resign

[Prashant Kafle @PrashantK7738]

Try it live: https://t.co/cFOjlIc7ML

Code: https://t.co/UDzHz4NmXa

Would love feedback from anyone working in RAG/LLM systems.

#AIEngineering #RAG #BuildInPublic

[Prashant Kafle @PrashantK7738]

I'm a CE student building toward remote AI Engineer / Forward Deployed Engineer roles. This is how I'm learning end-to-end — not just calling an API, but designing the retrieval architecture and product around it.
Open to opportunities 👋