Rishi Tiwari

Pulse (AI Observability and Evaluation Platform)

Unified release-quality platform for AI agents covering observability, simulation, and offline evaluation.

• What: One system for inline telemetry, simulation, and post-hoc evaluations across 20K+ daily conversations
• How: Temporal-orchestrated batch runs with mix-and-match model routing across OpenAI and Gemini
• Impact: Reduced recurring post-hoc evaluation spend by ~50%; platform capacity supports 1M+ evals/day (cost-governed in production)

Destination Content Store (MakeMyTrip)

Data generation and retrieval pipeline that powers destination-related queries on Myra.

• What: Automated content generation for destinations and attractions
• How: Indexed embeddings in Qdrant and modeled entity relationships in a knowledge graph
• Impact: Moved destination answers from generic LLM memory to DCS-powered retrieval with 30% higher groundedness and broader entity coverage (4-5 -> 10+)

Myra (MakeMyTrip)

First LLM-driven use case launched at MMT: a virtual flight booking assistant for chat and voice journeys in English and Hindi.

• What: Shipped core assistant flow from search through payment
• How: Productionized custom LangChain modules for low-latency execution
• Impact: Piloted to 1% of users while reducing production response time by ~90% (25-30 s -> 2-3 s)

Data Migration and Marketing ETL Pipelines

Data platform workstream focused on large-scale OLTP -> Delta Lake (OLAP) migration and downstream activation pipelines.

• What: Migrated critical OLTP analytics sources (including payments) to Delta Lake for OLAP consumption
• How: Built config-driven Spark pipelines tuned for multi-TB/day processing with minimal source-system impact
• Impact: Stabilized MMT analytics backbone while reducing setup time from 2 days to 3-4 hours; full dump jobs completed in <1 hour

Curie (Energy NL-to-SQL Platform)

Natural-language analytics workflow for enterprise energy management operations.

• What: Delivered NL-to-SQL access for enterprise operations teams
• How: Built conversational querying and alerting widgets into the web app
• Impact: Increased supported query types by ~80% in fewer release cycles

Reliability

• Design for graceful degradation under partial failures.
• Use clear ownership boundaries and observable runbooks.

Performance

• Treat latency budgets as product requirements.
• Profile bottlenecks first, then optimize with measurable deltas.

LLM Correctness

• Constrain prompts with task-scoped context and guardrails.
• Validate outputs with deterministic checks before user impact.

Backend Systems

LLM / AI

Data

Infrastructure