Saswatika Joshi

Owned architecture, engineering, and production deployment of enterprise generative AI solutions including agentic workflows, RAG, and LLM fine-tuning. Built multi-agent orchestration using LangChain and LlamaIndex, implemented tool registries and memory systems, and designed LLM smart routing with intent classification, fallback chains, and arbitration for reliability, latency, and cost. Developed end-to-end RAG pipelines from ingestion and chunking through embedding tuning, hybrid retrieval, reranking, multi-hop retrieval, and context management. Implemented guardrails, hallucination detection, and output validation; built model-serving infrastructure with inference endpoints/APIs, caching, observability, and resilient fallback logic. Automated testing/validation/deployment with GitLab CI/CD and supported model promotion controls and evaluation workflows.

Architected end-to-end ML pipelines covering ingestion, feature preparation, training, registry, lineage, governance, and reproducibility for enterprise deployments. Developed classification, forecasting, and anomaly-detection models using PyTorch, TensorFlow, and classical ML. Designed scalable hosting APIs and microservices with OAuth 2.0, retry logic, rate limiting, and production error handling. Hardened SQL and PySpark ETL/ELT pipelines using schema validation and data contracts; improved release reliability through Docker, Kubernetes, CI/CD, and observability.

Developed an AI assistant/voice bot solution using NLP/NLU and conversational AI techniques, with iterative evaluation using structured metrics. Applied audio signal processing (MFCC) and data augmentation to improve robustness across multimodal inputs. Built scalable speech-to-text and text-to-speech pipelines with sentiment classification, including supporting documentation of architectural decisions for team collaboration.