Principal Data Scientist | LLM / RAG / Conversational AI

📍 Bangalore, India
📧 [email protected]
🔗 LinkedIn
🔗 YouTube

Principal Data Scientist with 16+ years of experience (9+ in ML/AI) specializing in:

• LLM-powered conversational systems
• Retrieval-Augmented Generation (RAG)
• Multi-turn reasoning & ambiguity resolution

Built production AI systems serving:

• 👥 ~148K users
• ⚡ ~4K queries/day
• 📚 ~5K internal knowledge documents

👉 Improved system deflection from 26% → ~50%

Enterprise-scale conversational AI system designed to handle ambiguous user queries using LLMs, RAG, multi-turn reasoning, and retrieval optimization.

This system powers an AI Service Desk (AISD) used by ~148K employees, handling:

• ⚡ ~4,000 queries/day
• 📚 ~5,000 internal knowledge documents

👉 Improved automation/deflection from 26% → ~50%

Users frequently submit vague and ambiguous queries, such as:

• “I need help”
• “Unable to login”
• “Reset password”

• Missing context in user queries
• Over-specific refined queries causing LLM failures
• Poor UX (long, unstructured responses)
• Scaling ambiguity resolution across topics

User Query ↓ Query Understanding (Rephrase + Context) ↓ Semantic Retrieval (OpenSearch - Top 100) ↓ Reranking (Cohere - Top K) ↓ LLM Reasoning Layer ↓ Answer Generation ↓ Structured Response (UX Layer)

• Transitioned from single-turn → context-aware system
• Maintains conversation history for better reasoning

• Decision-tree-based disambiguation
• Handles structured flows for known topics
• Example flow:
  • Reset password → Ask OS (Windows/Mac) → Ask login status → Provide resolution

👉 Improved deflection: 26% → 40%

• Handles unconfigured / scalable scenarios
• Uses top-k retrieved documents to generate follow-ups

Problem:
Refined queries became too specific, causing LLM to fail (~10%)

Solution:

• Generate multiple simplified queries
• Remove unnecessary constraints
• Retry retrieval + generation

👉 Results:

• +10% answer coverage
• +5–6% deflection

Problem:

• Long LLM responses reduce readability in Slack

Solution:

• Structured output:
  • 🚀 Quick Solution
  • Step-by-step instructions
  • Notes / context

👉 Impact:

• +2–3% deflection
• Improved user engagement

• LLMs: LLaMA (llama4.scout)
• Architecture: RAG, Multi-turn reasoning
• Search: OpenSearch (semantic retrieval)
• Ranking: Cohere rerank
• Cloud: Oracle Cloud Infrastructure (OCI)
• Data: ATP DB
• Languages: Python, SQL

• Ambiguity handling is critical in real-world AI systems
• Retrieval quality directly impacts LLM performance
• Over-specific queries can degrade answerability
• UX design is as important as model performance

• Adaptive retrieval strategies (dynamic top-k)
• Better query intent classification
• Reinforcement learning from user feedback
• Latency optimization for real-time interaction

• 🥈 Ericsson ML Challenge — 2nd / 4120
• 🥉 USAID Forecasting Challenge — 3rd
• 🥉 MakeMyTrip — 3rd / 3556
• 🏆 BrainWaves (SocGen) — Finalist (4926)

• Building production-grade LLM systems
• Solving ambiguity in real-world AI systems
• Improving retrieval + reasoning + UX together

If you're working on:

• LLM systems
• RAG architectures
• Conversational AI

Feel free to connect or collaborate!