ML Engineer III

Vishak
Bharadwaj S

ML Engineer with 6+ years building data-driven systems across recommendation, deep learning, GenAI, and MLOps — from signal pipelines and ranking models to LLM-powered annotation and production monitoring at scale, delivering measurable engagement lifts and production-grade systems serving 100M+ users.

Get in touch Download CV LinkedIn ↗
scroll
01 — Experience

Work History

Machine Learning Engineer II → III
Jun 2022 — Present
Glance · Bengaluru, India
  • Hybrid Recommendation Engine: Designed and deployed a dual-track ranking pipeline that dynamically routes users based on data density — owning the Samsung channel (40M users) within a 150M-user platform. Powered the "dense" track with Gemini-enriched content embeddings, a Two-Tower retrieval model, and an LGBM ranker, while gracefully handling cold-start "sparse" users using Wilson's lower bound popularity and recency signals. Achieved a 40% lift in interactions — notable for a locked-in lock screen product where users never actively opt in.
  • Signal Extraction & Data Pipeline: Iterated extensively on high-volume, noisy event logs (150M+ user signals) to isolate clean, high-intent engagement signals, ensuring downstream ML models were trained on accurate behavioral data. 30-min hourly batch pipelines; sub-250ms online serving latency.
  • Infrastructure (Vertex AI → GKE): Built online and offline serving pipelines on Vertex AI with Vertex Feature Store; migrated to GKE + Argo CD for cost efficiency; wrote Golang prediction services and model controllers; instrumented with OpenTelemetry and Grafana.
  • Experimentation (Alchemist): Contributed to an internal A/B testing platform; derived minimum sample sizes from inference/confidence equations to ensure statistical significance before shipping ranking changes.
  • AI Annotation Setup: POC for annotation workflows; GenAI-based image metadata tagging with LLMs and prompt engineering for category classification; drove annotation cleanup and cost reduction.
Machine Learning Engineer I
Nov 2020 — Jun 2022
Censius AI · Bangalore, India
  • Explainability Module: Used SHAP and LIME to explain model predictions and provide insight into why models produce the outputs they do; logged and monitored using MLflow, Prometheus, WhyLogs and Grafana.
  • Drift Prediction Module: Created APIs for continuous monitoring of model performance and production data; monitored data and concept drift with custom code and WhyLogs; built data quality, drift and performance monitors on Prefect / Airflow jobs.
  • Deployment & Infrastructure: Containerized and deployed using Docker, GitHub Actions and AWS ECS; orchestrated workloads on Kubernetes.
  • ML Lifecycle: Worked across the full ML lifecycle — from model onboarding and deployment to post-production monitoring — enabling clients to detect model degradation early and act on it.
Machine Learning Intern → Jr Machine Learning Engineer
Jul 2018 — Oct 2020
Omni-Eye / The Valley Edutech · Bangalore, Karnataka
  • Built models for security systems for the Omni-Eye platform.
  • Eye In the Sky — Real-time Image Processing: uses stacked deep learning models to process images for object detection, facial recognition and plate detection. MTCNN network for facial detection; finetuned PyTorch models for the facial recognition aspect. OCR and data pipelines for object detection and plate recognition. Tracked project progress with MLflow.
  • Image Search & Clustering: Created, trained and finetuned a CNN autoencoder that converts unlabelled images into feature vectors; inserted into KNN and LSH (Locality Sensitive Hashing) for fast similarity search and into unsupervised clustering algorithms to group image data.
  • Student Platform & Instruction: Developed a portal for tracking student progress (Flask, MongoDB) with GitHub commit-tracking APIs; instructed Python, ML and Deep Learning cohorts with a code-first, project-oriented approach.
02 — Projects

Notable Work

ERAv4 · AWS EC2
ResNet50 on ImageNet-1k from Scratch
Trained ResNet50 without pretrained weights on full ImageNet-1k on AWS EC2 — completed by ~10,000 people globally. Achieved 75%+ top-1 accuracy. Deployed as a live demo on HuggingFace Spaces.
Chrome Extension
YouSum — AI YouTube Summarizer
Chrome extension that extracts transcripts from YouTube videos and generates streaming summaries via Claude and ChatGPT APIs. Supports 5 detail levels, background generation, persistent storage, and markdown rendering. GitHub ↗
Andrew Ng · C4
YOLO Object Detection
Implemented the YOLO algorithm for real-time object detection on an autonomous driving dataset — bounding box prediction, IoU, and non-max suppression from scratch.
Andrew Ng · C4
Face Recognition with FaceNet
One-shot face verification system using the FaceNet architecture and triplet loss.
Flask · Gemini AI
Poetry Analysis Studio
Flask web app that uses Google Gemini to analyse submitted poems and generate examples across 4 forms (Haiku, Sonnet, Limerick, Free Verse) with detailed literary analysis. GitHub ↗
Andrew Ng · C5
Neural Machine Translation with Attention
Seq2seq model with an attention mechanism for translation, learning to focus on relevant input positions at each decoding step.
Andrew Ng · C5
Trigger Word Detection
Constructed a custom speech dataset and implemented wake-word detection using a GRU-based model in TensorFlow and pydub.
Kaggle · Top 0.4%
Rossmann Store Sales Prediction
Deep network of embedding layers achieving 10% RMSPE. Ranked 11th out of 3,000+ teams.
Kaggle · Top 0.4%
BlueBook for Bulldozers
Random Forest Regressor achieving RMSLE of 0.2214. Ranked 2nd out of 476 teams — second-lowest error in the competition.
03 — Skills

Toolkit

Domains
Machine Learning Recommendation Systems Deep Learning MLOps Model Monitoring Computer Vision NLP Data Science
Stack
Python Go PyTorch Scikit-learn Pandas · NumPy PySpark Vertex AI GKE OpenTelemetry Grafana SHAP / LIME
04 — Background

Education & Certifications

Bachelor of Engineering
B M S College of Engineering · Bangalore
2012 — 2016
🎓
Deep Learning Specialization
Andrew Ng · deeplearning.ai / Coursera · 5 courses
  • C1 Neural Networks and Deep Learning
  • C2 Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization
  • C3 Structuring Machine Learning Projects
  • C4 Convolutional Neural Networks
  • C5 Sequence Models
🧠
END2 — Extensive NLP via Deep Models
The School of AI
  • NLP Transformers, attention mechanisms, BERT, GPT-1/2/3
  • NLP PyTorch for NLP, embeddings, and modern language modelling
  • NLP Retrieval-augmented generation techniques
🚀
ERAv4 — Extensive & Reimagined AI Program
The School of AI
  • LLM End-to-end LLM pretraining and instruction tuning from scratch
  • LLM Quantization-Aware Training (QAT), RLHF, and Vision-Language Models (CLIP)
  • LLM Multi-GPU CNN training on ImageNet; transformer architecture deep-dives
05 — Contact

Get In Touch

Email
vishak.svec@gmail.com
Phone
+91 948 362 8282
LinkedIn
linkedin.com/in/vishakbharadwaj
Location
Bengaluru, KA 560079