Sourabh Sharma

Senior Machine Learning Engineer (LLMs, ML Infra, Vision)

5+ years building production ML systems in e-commerce and B2B SaaS. I own the full stack from problem framing and data to type-safe pipelines, models, and experiments in production. I care about robust, observable ML infrastructure that moves product metrics.

Areas of Interest

LLMs
Agentic
GenAI
Computer Vision
Full Stack

Key Skills

Machine Learning & AI

PyTorch
TensorFlow
SageMaker
Vertex AI
Kubeflow
Weights & Biases
MLflow
DVC
ONNX
Optuna
Hydra
PyTorch Lightning
scikit-learn
DeepSpeed
XGBoost

Cloud & Infrastructure

Docker
AWS
Google Cloud
Cloudflare
Kubernetes
Terraform
Airflow
Prefect
OpenTelemetry

Data Engineering & Analytics

Dask
Spark
Apache Kafka
dbt
PostgreSQL
Redis
Elasticsearch
Apache Superset
BigQuery
Google Analytics
Ray
Apache Beam

Web Development & Design

React
Next.js
Tailwind CSS
FastAPI
Flask
Hono
Playwright
Selenium
Zustand
Zod
Drizzle ORM
Deno
tRPC

Programming & Development

Python
TypeScript
Bash
Go

Work Experience

Aidaptive by JarvisML

Cupertino, USA

AI-driven personalization platform focusing on hospitality and e-commerce.

Senior ML Engineer

Sep '22 – Present
Dynamic Pricing System - Twiddy VRM
  • Architected an ML-based dynamic pricing engine managing $475M+ in bookings for 1,240 vacation rental properties, with experiments indicating $5M+ annual revenue upside from optimized pricing.

  • Modeled demand and price elasticity with XGBoost over 60+ engineered features (GA4 demand signals, image-derived attributes, seasonal/geo elasticity, occupancy and booking lags), with features materialized in BigQuery.

Python
XGBoost
MLflow
BigQuery
Feature Engineering
Image Attribute Extraction - Cross-Client
  • Achieved F1 ≈ 0.9 on 100+ hierarchical attributes by batch-processing 200K+ images across 8 clients.

  • Integrated Gemini Vision and Claude for zero-shot attribute extraction.

PyTorch
Gemini Vision
Claude API
Vertex AI
Zero-shot Learning
ML Infrastructure Library (ml-lib)
  • 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.

Python
Vertex AI
BigQuery
Model Registry
NL→SQL Service "Ask a Metric"
  • Shipped an NL→SQL service with a Faiss vector store and result cache; reduced median response latency by matching queries and cut LLM costs 90%.

Faiss
RAG
FastAPI
LLaMA
GPT-3.5
SQL-Coder
  • Developed a visual similarity recommender that increased RPI by 14% for select clients by surfacing higher-value alternative products.

  • Built an internal GenAI experimentation framework using LoRA-based fine-tuning (LLMs / SDXL) to generate ad copy and creatives with consistent characters.

Subway India

Bangalore, India

Quick-service restaurant franchise with 1,000+ stores across India.

Independent Contractor

contract
Jan '25 – Mar '25
  • Delivered enterprise support platform via WhatsApp for Subway India with 1,065 stores (81% adoption) across India, processing 11K tickets (362K+ messages) with 56% sub-2-minute resolution (51-min median) through decision tree routing across 13 departments, laying foundation for ML-powered ticket routing.

  • Architected platform serving both franchise support and employee innovation (Sankalp) with HR/committee governance through 4-tier RBAC, event-driven state machines, JWT/OTP/SSO auth, WhatsApp Cloud API integration, serverless deployment with PostgreSQL and Redis.

Next.js
TypeScript
PostgreSQL
Redis
WhatsApp API
Vercel

Aisle3

London, UK

Early-stage e-commerce aggregator for multi-merchant product discovery.

Founding ML Engineer

Nov '20 – Sep '22
Multi-modal Product Matching Engine
  • First ML hire; owned the end-to-end design, training, and deployment of a multi-modal product matching engine (text + image) for noisy, long-tail e-commerce catalogs using Swin Transformers and self-supervised pretraining (DINO, SimCLR, MoCo), reaching P@1 ~0.9, Recall ~0.96 on a curated hard evaluation set.

  • Built a serverless, parallel, fault-tolerant workflow on AWS Step Functions (training → vector indexing → clustering → product groups), reducing infrastructure costs by ~60%.

  • Built a web-based active learning annotation tool that generated 120K+ pairwise labels via human-in-the-loop review and targeted sampling, significantly accelerating training data collection.

PyTorch
Swin Transformers
DINO
SimCLR
MoCo
AWS Step Functions
Active Learning
Online Vector Search Platform
  • Deployed a Faiss + Elasticsearch vector search platform (orchestrated with Airflow) processing 60GB+ / day, cutting duplicate listings by ~90% and powering features like product matching, color detection, and image-attribute extraction.

Faiss
Elasticsearch
Apache Airflow
Docker

Early Work(May '19 – Aug '20)

ML Engineer

@ FaceX
contract
5 mosBangalore, India

Startup providing face recognition and liveness detection solutions.

  • Built masked/peri-ocular face recognition for masked faces and a real-time face liveness SDK (client SDKs for mobile and backend services) used in Aadhaar e-KYC and fraud-prevention flows, handling 100K+ verifications/day.

Computer Vision
Face Recognition
Python
OpenCV

ML Engineer

@ Jumper.ai
contract
3 mosSingapore

Platform enabling social commerce through chatbots and automation.

  • Built NLP-based conversational AI flows for a social commerce chatbot platform by integrating backend APIs and a menu-digitization model for automated restaurant menu parsing, using ML-based intent and NER models.

NLP
Conversational AI
Python

Data Engineer

@ 73 Strings
contract
3 mosBangalore, India

Financial technology firm providing AI-driven valuation tools.

  • Developed ML- and rule-based crawlers and pipelines to scrape and structure company information, powering richer company profiles for downstream pipelines.

Python
Web Scraping
Data Pipelines

Intern

@ Gumption Labs
internship
2 mosBangalore, India

Startup focused on automated trading solutions for retail investors.

  • Experimented with genetic programming to explore trading rules and financial technical analysis using Selenium for automation.

Python
Genetic Programming
Selenium