Skip to content

Available for consulting engagements

Mukul Kumar · Computer Vision & Document AI Consultant

Production computer vision and document AI — accurate, fast, and cheap to run.

I help product and engineering teams ship ML systems that are accurate, fast at inference and cheap to run — 8 years taking CV and document AI to production, 5 patents filed.

Book an intro call See case studies

About

Hi, I'm Mukul.

I help product and engineering teams turn computer vision and document AI ideas into production systems. Over 8 years I've taken deep learning from research to production - most recently automating vehicle damage assessment for insurance claims, and before that, extracting structured data from scanned receipts in a system deployed across multiple countries.

I work across the whole lifecycle: framing the problem, building and evaluating models on your data, and making inference fit your latency and cost budget with TensorRT, Triton and DeepStream. After launch I set up the unglamorous parts that keep models useful - clean data pipelines, honest evaluation, drift monitoring and retraining.

My research background - 5 filed patents and 2 peer-reviewed papers, including a lightweight graph-attention text line detector presented at the KDD Document Intelligence Workshop - means I can bring current methods to your problem when they help, and skip them when a simpler model will do.

Focus
Computer vision · Document AI · Inference optimization
Based in
Bengaluru, India — working with teams in the US, EU and India, async-friendly
Engagements
[TODO: e.g. Fixed-scope audits, project builds, fractional ML lead]

  • 45%

    Lower manual inspection costs

  • 30%

    Lower annotation costs per year

  • 100–200×

    Smaller model at state-of-the-art accuracy

  • 8

    Years shipping ML to production

Services

How I can help

Inference optimization audit

Who it's for

Teams whose models work but are too slow or too expensive to serve - high GPU bills, missed latency targets, or a model that doesn't fit the target device.

What you get

  • Profiling of your serving path, from pre-processing to GPU utilisation
  • Optimization plan and prototypes with TensorRT, ONNX and Triton
  • Benchmark report: latency, throughput and cost per inference, before and after

TensorRT-optimised models shipped in insurance, document and video-analytics systems.

Typical timeline · [TODO: e.g. 2–3 weeks]

Discuss this - Inference optimization audit

Document AI & OCR pipelines

Who it's for

Companies processing receipts, invoices or forms at volume who need structured data out of messy scans and photos.

What you get

  • Pipeline design: text detection, OCR, line grouping and entity extraction
  • An evaluation set and field-level accuracy metrics on your own documents
  • A production-ready service with confidence scores for human review

Backed by 4 filed patents and a KDD workshop paper on receipt extraction.

Typical timeline · [TODO: e.g. 6–10 weeks]

Discuss this - Document AI & OCR pipelines

Computer vision: PoC to production

Who it's for

Product teams with an inspection, damage-detection or video-analytics idea - or a prototype that hasn't made it to production yet.

What you get

  • Feasibility check and a data and labelling strategy, including self-supervised pre-training
  • Model development with honest evaluation on held-out data
  • Deployment to cloud, edge or live video streams, with monitoring

Vehicle damage assessment for insurance and real-time analytics on 40+ CCTV cameras.

Typical timeline · [TODO: e.g. 8–12 weeks]

Discuss this - Computer vision: PoC to production

ML system review / fractional ML lead

Who it's for

Engineering and data leaders who need senior ML judgement without a full-time hire - before a big build, or when production models start to drift.

What you get

  • Review of data pipelines, evaluation and MLOps, with prioritised findings
  • A drift-monitoring and retraining strategy
  • Ongoing technical leadership: roadmap, model reviews and hiring support

Typical timeline · [TODO: e.g. 2-week review, or N days per month]

Discuss this - ML system review / fractional ML lead

Selected work

Case studies

2023 – 2026

Automated Vehicle Damage Assessment

Ensemble computer-vision system that detects vehicle damage and missing parts for insurance claims - 90% damage-detection accuracy.

  • Computer Vision
  • Self-Supervised
  • PyTorch
  • TensorRT
  • AzureML
Read case study

2019 – 2023

Information Extraction from Receipts

CV + NLP pipeline that extracts and decodes purchased items from scanned receipts. Deployed in multiple countries; 4 patents and a KDD workshop paper.

  • Document AI
  • GNN
  • NLP
  • OCR
  • DGL
  • TensorRT
Read case study

2020

Social Distancing Monitoring System

Real-time, multi-camera system that flags social-distancing violations from existing CCTV feeds.

  • Computer Vision
  • DeepStream
  • TensorRT
  • Video Analytics
Read case study

Testimonials

What people I've worked with say

He collaborated and communicated well as part of a diverse, international and multidisciplinary team and he participated in key innovations also pushing the boundaries of the state of the art.

Read full recommendation

Research and innovation requires thoughtful data analysis, continuous learning and autonomy as demonstrated by Mukul. During the three years he worked on my team he grew as a professional specializing in CV, DL, ML and GNNs in the boundaries of image and text processing. He collaborated and communicated well as part of a diverse, international and multidisciplinary team and he participated in key innovations also pushing the boundaries of the state of the art. Mukul would be a valuable asset to any team in the area of ML, willing to adapt and tackle new challenges.

Javier Yebes

Director, AI Research Science, NielsenIQ

Managed Mukul Directly

Recommended on LinkedIn(opens in a new tab)

How I work

From first call to hand-over

  1. Discover

    Understand the business goal, the data and the constraints - accuracy, latency, cost. Agree on success metrics up front.

  2. Build

    Prototype quickly, evaluate honestly on your own data, and iterate toward a model worth shipping.

  3. Optimize

    Make it fast and cheap to run - TensorRT, ONNX, Triton, batching and quantization where they pay off.

  4. Hand over

    Documentation, monitoring and retraining playbooks, and knowledge transfer so your team owns it.

How I deliver

Tools for each stage

Research

  • Transformers & GPT
  • CNNs & GNNs
  • Diffusion & GANs
  • Self- & weakly-supervised learning

Build

  • PyTorch
  • TensorFlow / Keras
  • Hugging Face
  • Deep Graph Library
  • OpenCV
  • spaCy
  • Python
  • C / C++

Optimize

  • TensorRT
  • ONNX
  • Triton Inference Server
  • NVIDIA DeepStream
  • PyCUDA

Operate

  • Docker
  • AzureML / MLflow
  • TF Serving
  • AWS / GCP
  • Drift monitoring & retraining

Research & patents

Publications, patents and research projects

Peer-reviewed papers

Patents filed

  1. Methods, Systems, Apparatus and Articles of Manufacture to Detect Lines on Documents

    Filed Jun 2022 · US81259087

  2. Automatic Document Content Extraction and Decoding

    Filed Jun 2021 · US81254014

  3. Automated Extraction of Purchased Items in Images of Receipts Using Computer Vision and Deep Learning

    Filed Jun 2021

  4. Automated Receipt Decoding: Product Matching and Dictionaries

    Filed Jun 2021

  5. Method and System to Detect Distance Between Entities

    Filed Aug 2020 · IN 202021034381

Research projects

2022

GAT for Text Line Detection

4M-parameter Graph Attention + GRU model for line detection on receipts - state-of-the-art 97.23% F1 on CORD with 100–200× fewer parameters.

  • Research
  • Graph Attention Networks
  • Document AI
Read more

2023

Parkinson's Detection from Voice

Vocal Tab Transformer that classifies Parkinson's disease from dysphonia features, beating gradient-boosted trees by 1%+ AUC.

  • Research
  • Transformers
  • Healthcare
  • Tabular
Read more

Background

Where I've done this before

Experience

  1. XA Group (Xpress Automation) · Senior Data Scientist

    2023 – 2026

    Led deep learning for automated motor-insurance claims - 45% lower manual inspection costs.

  2. Blackstraw.ai · Senior Data Scientist

    2019 – 2023

    Consulted with client teams; lead developer on receipt extraction deployed in multiple countries, real-time CCTV analytics - 5 patents filed.

Education

  1. M.Tech, Artificial Intelligence & Machine Learning

    Birla Institute of Technology and Science (BITS), Pilani - WILP

    2024 - 2026

  2. B.Tech, Computer Science & Engineering

    College of Engineering Roorkee

    2014 - 2018

Side projects

Built on the side

LEHO

leho.in

Website for the Ladakh Environment and Health Organization, a Leh-based NGO working on community development and climate adaptation

(opens in a new tab)

Voyager Ladakh

voyagerladakh.com

Website for a Leh-based travel company running treks, expeditions and motorbike tours across Ladakh.

(opens in a new tab)

Journal

Life beyond the code

ladakh cover image

· 1 min read · Leh, Ladakh, India

Ladakh experience

My journey and learnings.

  • Travel
Read all posts

Contact

Let's build something together

Have a computer vision, document AI or inference problem? Tell me about the project, timeline and constraints - I'll reply within [TODO: e.g. 2 business days].

Or email me directly at contact@mukulkumar.dev.