Ayushi Tiwari

Conversational AI System at Scale

• Designed and developed large-scale conversational AI systems integrating APIs across 1000+ hospitals (AIIMS, Telangana, Tamil Nadu, Railways), enabling patients to access lab reports, prescriptions, OPD schedules, payments, and test tariffs.
• Optimized Rasa NLU pipelines with 50+ intents and entity handling for noisy, unstructured, misspelled, and context-poor user queries in real-world healthcare scenarios, enhancing user interaction and query resolution.
• Engineered a multi-stage fallback pipeline:
  Rasa NLU → LLM-based query rephrasing & spell correction → reprocessing → semantic FAQ retrieval
• Integrated LLaMA-based prompting to dynamically rephrase user queries, significantly improving response success rates for ambiguous and incorrect inputs.
• Implemented multilingual support for Hindi, English, Tamil, Telugu, Punjabi, including:Transliteration handling (Hindi written in English script)
  Script detection (Devanagari vs Latin)
  Dynamic translation pipelines before NLU processing

Semantic FAQ Retrieval (FAISS + Embeddings)

• Implemented semantic FAQ retrieval using Sentence Transformers and FAISS, achieving context-aware matching of user queries with hospital knowledge bases, resulting in increased response relevance.
• Improved response accuracy by replacing keyword-based matching with embedding-based similarity search.

Optimized entity mapping process to ensure accurate term recognition.

• Replaced manual entity-to-keyword mapping with embedding-based semantic normalization, enabling automatic mapping of user inputs (e.g., “CBC”) to hospital-specific terms (e.g., “Complete Blood Count”).
• Reduced manual configuration effort by ~70% and improved scalability across multiple hospital systems.

LLM-Based Structured Extraction

• Designed LLM-driven pipelines using LLaMA to convert doctor–patient conversational text into structured JSON outputs (complaints, vitals, medicines, diagnosis, etc.).

Performance & Reliability Engineering

• Reduced AI response latency by ~5 seconds (~60% improvement) by implementing model pre-loading instead of per-request loading.
• Implemented timeout handling, retry mechanisms, and fallback strategies for external API failures, ensuring system stability and reliability.

Code Architecture & Deployment

• Modularized large and evolving codebases to support scaling across multiple state-level hospital systems, improving maintainability and code reuse.
• Managed UAT and production Linux environments, including deployments, log monitoring, debugging, and issue resolution.