Karan B.

I am an AI/ML Engineer specializing in production-ready LLM, RAG, NLP, and machine learning systems. I build scalable AI applications using Python, PyTorch, Transformers, FastAPI, LangChain, Qdrant, Docker, Kubernetes, and MLOps/CI/CD workflows. I have experience building production chatbots, RAG assistants, agent benchmarking systems, data generation pipelines, synthetic datasets, recommendation engines, and end-to-end ML workflows from data preparation to deployment. I can help with LLM and RAG application development, AI chatbot development, NLP model training and fine-tuning, vector database integration, FastAPI-based AI/ML backend development, Docker and Kubernetes deployment, MLOps and CI/CD pipelines, agent evaluation and benchmarking, recommendation systems, ranking models, model evaluation, monitoring, and production readiness. In my current Applied AI Engineer role at Abundant (YC), I partner with leading AI research labs under NDA to support production-grade agent benchmarking and evaluation. I build evaluation and data generation pipelines, curate synthetic datasets, ship Docker and Kubernetes-based test environments, analyze agent behavior and failure modes, adapt open-source research repositories, and implement GitHub workflow checks, health monitoring, and data integrity tests for reliable and reproducible large-scale experiments. Previously, as an AI/ML Engineer at Fusemachines, I built and production-hardened a global AI Fellowship chatbot using FastAPI, RAG, Qdrant, and Groq LLMs. I implemented semantic in-memory caching with 12,000 entries, dual-LLM failover, resiliency controls, rate limiting, circuit breakers, Prometheus metrics, and Kubernetes health probes. The system achieved approximately 800ms average latency, under 850ms P95 latency, and a 28 to 32 percent cache hit rate. I have also built applied AI projects such as Sambodhan, an AI-powered grievance redressal platform for local governance, and RecipeShare, an AI-powered recipe recommendation platform. Sambodhan achieved over 95 percent complaint classification accuracy using XLM-RoBERTa, urgency detection, automated routing, and a production RAG assistant powered by LLaMA 3.3 70B through Groq. RecipeShare used Spring Boot, FastAPI, React, PostgreSQL, Kafka, Redis, Elasticsearch, ALS collaborative filtering, and a PyTorch Two-Tower model trained on more than 700,000 user interactions. My goal is to help clients build AI systems that are reliable, scalable, measurable, and ready for real-world use.