- The problems you will face multi-objective ranking across 6 platforms with competing commercial and user goals, demand forecasting for Indian festival travel with structural breaks, NLP that understands Hinglish booking intent, and LLM agents that reliably execute booking workflows do not have off-the-shelf solutions.

- You will define the research agenda: what problems are worth solving, what methods are worth trying, what results are worth shipping.

- Your work will directly affect 100M+ users a 0.2% lift in search ranking converts to millions of additional bookings annually.

- You are expected to publish, to present, and to set the intellectual standard for the entire AI team.

- This is not a notebook role. Models you design will run in production pipelines serving millions of predictions per hour.

Domains of Ownership :

Search Ranking & Relevance :

- Design multi-stage ranking pipelines: candidate retrieval (ANN search, BM25) coarse ranking (GBDT) fine ranking (neural IR) re-ranking (business rules + ML)

- Multi-objective optimisation: joint ranking for user satisfaction (CTR, dwell time) + commercial value (GMV, margin) + diversity + freshness

- Position bias and presentation bias correction; counterfactual offline evaluation (IPS, SNIPS)

- Cross-platform ranking: shared feature representations across flights, hotels, bus, and train inventory

Price Intelligence & Demand Forecasting

- Fare prediction models: price trajectory forecasting, optimal booking window recommendations, price alert intelligence

- Demand forecasting with structural breaks: Indian holidays, festival seasons, IPL calendar, election schedules

- Price elasticity modelling per segment (leisure vs. business, tier-1 vs. tier-2 cities, device type)

- Lower-price guarantee: real-time competitive price monitoring, margin-aware trigger logic, reimbursement risk scoring

- Yield management: seat/room inventory optimisation, dynamic pricing recommendation to suppliers

NLP, LLM Research & Voice AI

- Query understanding: named entity recognition (destination, date, passenger type), intent classification, query expansion for sparse queries

- Hinglish NLP: code-mixed text normalisation, transliteration-aware tokenisation, cross-lingual transfer learning

- LLM fine-tuning research: LoRA/QLoRA strategies for travel domain adaptation, RLHF for booking dialogue quality

- RAG architecture: retrieval quality evaluation, chunking strategies for long-form fare documents, hybrid (dense + sparse) retrieval

- Voice AI research: ASR accuracy benchmarking on Indian accents, TTS naturalness evaluation, spoken intent understanding

Recommendation & Personalisation

- Cross-vertical user modelling: unified embedding space across flight, hotel, bus, train behaviour signals

- Session-based recommendation: transformer-based sequential models (SASRec, BERT4Rec) for real-time session context

- Cold-start research: content-based initialisation, bandit-based exploration for new users and new inventory

- Contextual personalisation: time-of-day, device, location, co-traveller type, booking horizon as context features

Causal Inference & Experimentation

- Experiment design: CUPED/CUPAC variance reduction, interleaving experiments for ranking, cluster-based designs for marketplace interventions

- Causal uplift modelling: heterogeneous treatment effects for coupon targeting, pricing interventions, UI experiments

- Long-term effect estimation: surrogate metrics, switchback designs for policy-level changes

What Makes This Technically Challenging :

Scale: your models process 500M+ events daily offline evaluation at this scale requires infrastructure, not pandas DataFrames

Complexity: ranking must simultaneously satisfy user intent, supplier contracts, inventory signals, and revenue goals in real-time

Novelty: Hinglish NLP and Indian travel demand patterns are under-researched you are building new knowledge, not applying existing papers

Causal hardness: separating the effect of a ranking change from seasonal demand shifts, competitor pricing, and UX changes simultaneously

LLM reliability: booking is a high-stakes action LLM agents must refuse gracefully, confirm correctly, and never hallucinate a price or availability

Voice AI frontier: real-time, low-latency spoken booking in 20+ Indian languages with background noise, accent variation, and network degradation

The AI-First Mandate :

- AI is not an enhancement. It is the product architecture. Every surface, every API, every decision point is either ML-powered today or on the roadmap to be.

- Search & Ranking - Learning-to-Rank across flights, hotels, bus routes, train coaches; real-time re-ranking on user signals

- Voice AI - Hindi/Hinglish voice booking, intent resolution, spoken fare comparisons, accessibility-first conversational UX

- RAG Systems - Fare rule retrieval, hotel cancellation policy Q&A, airline contract intelligence, real-time regulatory updates

Agentic AI - Autonomous booking resolution, exception handling, refund orchestration, supplier communication bots

- MCP Orchestration - Model Context Protocol tool chains across GDS APIs, payment gateways, and supplier integrations

- Recommendation Engine - Cross-vertical next-best-action, collaborative filtering, session-based deep learning

- Price Intelligence - Competitive fare mapping, lower-price guarantee engine, demand elasticity, yield optimisation

- Coupon & Promo ML - Personalised offer targeting, redemption probability scoring, margin-aware discount optimisation

- Sentiment & Review AI - Review summarisation, NPS prediction, complaint triage, trust signal extraction

- Fraud & Risk ML - Anomaly detection, account takeover signals, payment fraud scoring, fake review classification

- Deep System Mapping - Route intelligence, geo-semantic matching, multi-modal journey planning

Predictive Systems - Cancellation risk, no-show prediction, seat upgrade probability, waitlist conversion

Who You Are :

- 10 to 14 years in ML/AI; PhD or equivalent research depth in a quantitative field strongly preferred (CS, Statistics, Operations Research)

- Published research or equivalent production depth in ranking, NLP, recommendation, or causal inference you have done original work, not just applied papers

- Hands-on production experience: models you have designed are running in serving infrastructure, not just Jupyter notebooks

- Deep LLM expertise: you have fine-tuned models, evaluated RAG pipelines, and reasoned about hallucination mitigation in production

- Fluent in Python, PyTorch/JAX, and distributed data processing (Spark/Flink)

- Tier-I institute strongly preferred (IIT / IIIT / IISC / BITS CSE / AI / Statistics / EE)

Technology Stack :

ML/DL : PyTorch JAX XGBoost LightGBM Scikit-learn CatBoost

NLP/LLM : HuggingFace Transformers PEFT vLLM LangChain DSPy Instructor

Voice : Whisper IndicASR Bhashini Speechbrain

Experimentation : MLflow Weights & Biases Optuna Custom eval frameworks

Data at Scale : Spark Flink BigQuery dbt Delta Lake Feast