Abhik Sarkar's avatar
Home
Abhik Sarkar

Abhik Sarkar

Machine Learning Engineer

abhiksark@gmail.com
www.abhik.ai
Download CV

Impact at Scale

Key metrics from production ML systems

Production Systems
7M+
Videos/Day

Daily video processing scale

81/sec
Throughput

Real-time inference speed

6x
Speedup

TensorRT + NVDEC optimization

98%
Deterrence

Crime prevention accuracy

3→30
Team Growth

Built & scaled ML team

$96K
Saved/Year

Infrastructure optimization

Technical Achievements

🚀
Architecture

GPU-Accelerated ML Pipeline

Designed end-to-end video processing pipeline with TensorRT 10.9, NVDEC hardware decoding, and zero-copy GPU operations.

TensorRTCUDANVDECPythonC++
6x speedup81 clips/secZero-copy ops
🤖
ML Innovation

GenAI Synthetic Data Pipeline

Built human-in-the-loop annotation system with GenAI-driven synthetic data generation for training data augmentation.

Diffusion ModelsCVATPyTorchLabel Studio
10x efficiencyAutomated QAActive learning
🛡️
Computer Vision

Real-time Threat Detection

Multi-model ensemble for trespass, theft, and anomaly detection powering nationwide security deployments.

YOLOv8DeepSORTAction RecognitionAnomaly Detection
98% deterrenceReal-time<100ms latency
⚙️
Infrastructure

MLOps & Model Versioning

Automated retraining orchestration with reproducible experiments, A/B testing, and gradual rollouts.

KubernetesMLflowDVCAirflow
CI/CD for MLAuto-scalingVersion control

Professional Experience

Cloudastructure Inc logo

Director, Machine Learning

Cloudastructure Inc

Nov 2020 - Present

Remote / Bangalore / Salt Lake, India

Joined as ML Engineer, promoted to Director in Aug 2023
7M+
Daily video clips processed
81/sec
Processing speed
6x
Performance speedup
$96K
Annual cost savings
Solo architected and optimized GPU-accelerated ML Tagger Pipeline, scaling video processing from 100K to 7M daily clips (81 clips/second), achieving 6x speedup via TensorRT 10.9, NVDEC, and zero-copy operations, resulting in $96,000 annual cost savings.
Founded and scaled 30-person India Global Capability Center (GCC) for R&D, data labeling, ops, and QA as Additional Director, enabling cost-efficient global expansion and supporting nationwide MSA rollouts.
Built and led a high-performing ML team from 3 to 30 members, fostering expertise in deep learning, scalable AI solutions, and cross-functional collaboration with DevOps, Frontend, and Backend teams.
Designed and implemented robust model version control, automated retraining/orchestration, and human-in-the-loop AI-assisted annotation workflows, improving model accuracy and boosting human annotation efficiency by 10x via GenAI-driven synthetic data generation.
Led cloud-to-colocation migration, reducing infrastructure costs by 75% while maintaining uptime and data integrity; negotiated with stakeholders to offset operational costs.
Contributed to 98% crime deterrence rate in multifamily housing MSAs through real-time threat detection (trespass, theft, anomalies), powering nationwide rollouts.
Links:India GCC LaunchHuman-in-the-Loop AIIndia EntityHousing MSA Press
Quantiphi Analytics Solutions Pvt. Ltd. / AthenasOwl logo

Machine Learning Engineer

Quantiphi Analytics Solutions Pvt. Ltd. / AthenasOwl

April 2019 - Nov 2020
Worked as part of the Video Intelligence Team at Athenas Owl, a Media-Based AI Product company.
Contributed to the development of a cutting-edge product aimed at assisting marketers in cataloging sports moments from vast video content libraries, spanning thousands of hours. Utilized a complex pipeline for different types of classification, object detection, and Siamese Network.
Responsible for generating new AI features in the product.
Contributed as an Engineer to a cutting-edge project aimed at developing a comprehensive athlete tracking solution across different sporting categories for amajor global sporting event. Focused heavily on OCR and Tracking. Ensured strict adherence to GDPR (General Data Protection Regulation) guidelines throughout the development process, prioritizing the protection of athletes' privacy and data rights.
Deloitte USI Consulting logo

Business Technology Analyst

Deloitte USI Consulting

June 2018 - March 2019
Worked in Human Capital Service Line which deals with research, analysis and design of critical programs involving different aspects of HR Processes.
Got Trained in Workday which is a cloud-based ERP Solution for human capital management and financial management applications.
Staffed in worldwide implementation of Financial Giant wherein Delivered both Inbound and Outbound Solutions using Workday Studio and EIBs(Enterprise Interface Builder).

Technical Expertise

Languages

4 skills

Python
PythonLearn more →
5
C++
C++Learn more →
2
JavaScript
JavaScriptLearn more →
2
Shell
ShellLearn more →
3

Frameworks, Libraries and Tools

11 skills

FastAPI
FastAPILearn more →
2
Next.js
Next.jsLearn more →
2
Tailwind CSS
Tailwind CSSLearn more →
2
React
ReactLearn more →
4
Docker
DockerLearn more →
3
Kubernetes
KubernetesLearn more →
3
PyTorch
PyTorchLearn more →
4
TensorRT
TensorRTLearn more →
4
ONNX
ONNXLearn more →
3
MLflow
MLflowLearn more →
3
Label Studio
Label StudioLearn more →
3

Databases

7 skills

MongoDB
MongoDBLearn more →
3
PostgreSQL
PostgreSQLLearn more →
3
Redis
RedisLearn more →
3
Elasticsearch
ElasticsearchLearn more →
3
ClickHouse
ClickHouseLearn more →
3
Kafka
KafkaLearn more →
3
Qdrant
QdrantLearn more →
3

Concepts

7 skills

M
Machine Learning
4
D
Deep Learning
4
C
Computer VisionLearn more →
4
N
Natural Language Processing
3
Video Processing
Video Processing
4
Real-time Inference
Real-time Inference
4
Object Detection
Object Detection
4

Things I'm Learning

5 skills

Golang
GolangLearn more →
Rust
RustLearn more →
Distributed Systems
Distributed Systems
WASM
WASMLearn more →
Linux Internals
Linux InternalsLearn more →

Speaking & Talks

Rolling with Python: Intro to Python Wheels

BangPypers Meetup 2024, Bangalore

Introduction to Python Wheels, why it is important, how it helps in packaging and distributing Python libraries. How C/C++ libraries can be included in Python Wheels.

PythonPackagingWheels

Speeding up Python with Cython

Pycon India 2024, Bangalore / Pycon Japan 2024, Tokyo

Under Core Python: Basic of PVM, Cython, how to speed up Python code, how it helps in Preprocessing and Postprocessing of data with Object Detection.

PythonCythonPerformance
View all talks

Education & Certifications

National Institute of Technology, Raipur logo

Bachelor of Technology in Computer Science and Engineering

National Institute of Technology, Raipur

2014 - 2018

  • Thesis: "Diabetic Retinopathy Detection using Deep Learning"
  • Pre-Thesis: Deposist Prediction using Machine Learning Models
  • Finalist in the Smart India Hackathon 2018
  • Winner of the NIT Raipur Model Making 2017
Udacity logo

Data Analyst Nanodegree

Udacity

2018 - 2019

  • Completed the Data Analyst Nanodegree
  • Projects: Investigate a Dataset, Analyze A/B Test Results, Wrangle and Analyze Data, Communicate Data Findings
  • Skills: Python, Pandas, Numpy, Matplotlib, Seaborn, Jupyter Notebook
Stanford Center for Professional Development logo

CS 224w: Machine Learning with Graphs

Stanford Center for Professional Development

2021

  • Completed the course on Machine Learning with Graphs
  • Skills: Graph Neural Networks, Graph Convolutional Networks, GraphSAGE, Graph Attention Networks
CCE Indian Institute of Science, Bangalore logo

Introduction to High-Performance Computing

CCE Indian Institute of Science, Bangalore

Aug 2023 - Dec 2023

  • Single Semester Course on High-Performance Computing
  • Skills: MPI, OpenMP, CUDA, Parallel Programming

Open Source

labelimgplusplus

PyPI Package

Enhanced graphical annotation tool for ML projects with bounding box labeling, multi-format export (PASCAL VOC, YOLO, CreateML), gallery mode, and undo/redo support.

PythonComputer VisionAnnotationPyQt5
PyPIGitHub

Learning Resources

People I Learn From

George Hotz
George Hotz
Andrew Ng
Andrew Ng
Andrej Karpathy
Andrej Karpathy
Robert C. Martin
Robert C. Martin
Arpit Bhayani
Arpit Bhayani
Hussain Nasser
Hussain Nasser

YouTube Channels

3Blue1Brown
3Blue1Brown
Yannic Kilcher
Yannic Kilcher
Dave Xiang
Dave Xiang
Szymon Ozog
Szymon Ozog
Core Dumped
Core Dumped
Byte Byte Go
Byte Byte Go

Podcasts

The Joe Rogan Experience
The Joe Rogan Experience
Huberman Lab
Huberman Lab
The Lex Fridman Podcast
The Lex Fridman Podcast
Real Python
Real Python

Recommended Reading

BRRR: A faster Python

by Horace He

The Unreasonable Effectiveness of Transformer Models

by Andrej Karpathy

Transformers from Scratch

by Brandon Rohrer

View all bookmarks
Mastodon