Bitan Biswas

Designed and deployed enterprise-grade conversational AI agents using LLM fine-tuning (GPT, LLaMA), prompt engineering, and RAG pipelines, achieving 30% improvement in containment rates and 25% reduction in fallbacks. Architected Agentic AI infrastructure and MCP servers enabling autonomous agent execution; developed multi-agent systems (CrewAI, n8n, UV) for decentralized workflows, reducing manual intervention by 40%. Built Node.js API integrations with CRM (Salesforce) and ERP (SAP) systems for real-time transactional workflows; implemented multilingual NMT and privacy-compliant human handoff strategies for GDPR/CCPA compliance. Enhanced model performance through structured A/B testing, iterative prompt refinement, embedding optimization, and LoRA-based fine-tuning of LLaMA architectures for production deployment.

Architected real-time B2B pricing engine processing thousands of daily transactions, reducing order processing time by 25%; engineered Kafka-based workflow automation across 10+ pricing workflows with exactly-once semantics. Designed 15+ secure REST APIs (Spring Boot) reducing ETL overhead by 10%; achieved 85% test coverage using Spock/JUnit, reducing production incidents by 30%. Built LLM chatbot (GPT-4 + LlamaIndex) for ARIBA procurement, reducing manual ticket handling by 50%; implemented production RAG pipelines with vector databases (Pinecone, FAISS), improving answer relevance by 40% and reducing hallucinations by 60%. Fine-tuned T5/BERT for contract classification and summarization with evaluation pipelines (ROUGE, BLEU, perplexity); deployed cloud infrastructure (Kubernetes, Terraform) with observability stack (New Relic, OpenTelemetry, Grafana).

Contributed to NLP-profiler open-source package, enhancing grammar and spelling checks with fuzzy similarity scoring (Levenshtein, Jaro-Winkler), improving contextual accuracy by 15%. Optimized CI/CD pipeline using GitHub Actions and dependency pruning, reducing runtime by 8% and improving package stability and deployment reliability. Collaborated in international distributed team across 3 time zones, participating in design discussions, code reviews, and Agile development cycles. Implemented contextual grammar models and similarity algorithms using spaCy and NLTK, improving text correction accuracy by 12% in benchmark evaluations.