Sourabh Sharma

Built demand-driven pricing engine (R² 0.75) for $175M+ annual revenue (57K+ bookings, 1,240 properties); identified 8-10% revenue optimization potential validated through backtest.

Modeled demand and price elasticity with XGBoost (Optuna-tuned, 5-fold CV) over 67 engineered features (GA4 demand signals, image-derived attributes, seasonal/geo elasticity, occupancy and booking lags), with features materialized in BigQuery and calibrated outputs.

Deployed ML-based image reranking across 8.4K+ listings (49.4M impressions), achieving +172% CTR lift and 62% win rate on underperforming properties through visual attribute scoring.

Engineered hybrid attribute extraction: trained vision classifiers on 11M+ web-crawled images achieving macro-F1 0.9 on 100+ hierarchical attributes, augmented with Gemini Vision and Claude zero-shot labeling; serving 200K+ images for 8 enterprise clients.

Extended Flagr in Go for real-time inference (sub-100ms), deployed on Kubernetes with monitoring and telemetry.

Designed and shipped a company-wide ML infrastructure library and model registry (Vertex AI + BigQuery + internal services) that became the default backend for new ML features across vision and LLM workloads.

Architected NL-to-SQL system using OpenAI Agents SDK, enabling non-technical stakeholders to query business metrics in natural language, with self-serve analytics across 4 data domains and 59 hospitality clients ($5B+ revenue data).

Developed FastMCP server exposing BigQuery operations as tools (schema introspection, guarded query execution) for agentic reasoning with Gemini (SQL generation) and GPT-4.1 (response synthesis).

Evolved from self-hosted 7B model (defog/sqlcoder) to production pipeline with semantic query catalog retrieval, syntax-aware CTE-based chunking, feedback-driven scoring, and tiered caching.

First ML hire; built multi-modal product matching with 200M-param Swin Transformer (DINO/SimCLR/MoCo contrastive pretraining) on 200GB+ image-text pairs, achieving P@1 0.91, Recall 0.96; with XGBoost + cross-encoder re-ranker combining vision, text, and attribute signals.

Designed serverless pipeline on AWS Step Functions, separating GPU/CPU workloads for cost-effective scaling.

Built annotation platform collecting 120K+ pairwise labels via active learning and HITL with embedding-based hard-negative sampling for contrastive model training.

Deployed Faiss + Elasticsearch for concurrent dense vector search, enabling product matching at scale (60GB+/day).

Designed self-training data flywheel for color detection: rule-based (CIELAB/deltaCIEDE2000) → CNN (ConvNeXt) → ViT (SwinV2), each generation producing weak labels to bootstrap the next; U²Net segmentation for salient object detection.

Developed multi-headed attribute extractor (progressive training on shared backbone) for footwear attributes (pattern, material, ankle height), achieving F1 0.93; powered product filtering on the storefront.