About
The through-line has been the same question in different clothes: how do you know the model is right. At Samsung it was image restoration you could judge by eye. At Michelin it is a natural-language interface to a pricing warehouse, where the answer looks equally plausible whether or not it is correct, and eyeballing it stops working.
That is why most of what I build now is evaluation and guardrails rather than models. Benchmarks that exist before the system does, read-only enforcement that holds at every layer, gates that stop the agent when it is uncertain instead of letting it guess confidently.
I am in Pune, India, and I am looking for work where the system has to hold up under real users rather than a demo.
Positions
Michelin
Michelin
Michelin
Samsung R&D Institute India
Education
B.Tech, Computer Science (Big Data Analytics), GPA 9.25 / 10
Skills
Generative & Agentic AI: LangGraphLangChainReAct agentsMulti-agent systemsMCPRAGPrompt engineeringFine-tuning (LoRA, QLoRA)
Machine Learning & NLP: PythonPyTorchTensorFlowTransformersLLMsModel evaluationAnomaly detectionForecasting
ML Engineering: Azure MLMLflowDockerKubernetesCI/CDREST APIsSQLDremioReactTypeScript