Kapil Mirchandani | ML Engineer

About Me

I'm a Machine Learning Platform Engineer at Spotify, where I work on Agent Platform and Machine Learning infrastructure. I've spent the past few years building end-to-end ML systems across different domains, from NLP and computer vision to LLM-powered products, and I enjoy building scalable infrastructure and shipping ML systems to production.

Before Spotify, I shipped ML features at Ford, Avoma, and Helpshift, working on everything from transformer-based semantic similarity models and RAG pipelines to real-time inference services handling millions of daily requests. I'm proficient in Python, PyTorch, and TensorFlow, and have deep experience with MLOps — CI/CD pipelines, model serving, GPU provisioning and management, and performance optimization.

I hold a Master of Engineering in Applied AI from the University of Ottawa. I've also published two IEEE papers on Generative Adversarial Networks and Big Data Analytics.

Experience

Where I've worked and what I've built.

Spotify

Platform Engineer - Machine Learning Infrastructure

April 2025 – Present · Toronto, Canada

Building Agent platform and Machine Learning infrastructure at Spotify.
Led end-to-end evaluation and deployment of an LLM observability platform in a hybrid cloud architecture — from vendor selection (10 candidates) through production rollout with SSO, data migration, and user onboarding.
Designed multitenancy access model with automated user provisioning; built declarative permissions management and authored the platform release process.
Extended ML inference gateway to support audio modalities and additional model providers.
Conducted autoscaling experiments on GPU-backed LLM serving infrastructure (SGLang) to evaluate server metrics and identify the optimal signal for GPU autoscaling.

Natural Resources Canada

Data Scientist Intern — Quantum AI Research

January 2025 – April 2025 · Ottawa, Ontario, Canada

Built and evaluated QMLP and QCNN models for multiclass classification of satellite images using PennyLane with a PyTorch backend, achieving ~40% accuracy — strong for current Quantum ML limitations.
Developed classical neural networks with parameter constraints for comparative analysis with quantum models, conducting detailed benchmarking of benefits and trade-offs.

Ford Motor Company

Software Developer — SYNC HMI Intern

May 2024 – December 2024 · Ottawa, Ontario, Canada

Implemented an automated bug ticket triaging system using ML on Jira tickets, achieving 80% routing accuracy and saving 15 developer hours per week — reducing tickets requiring manual triage by 50%.
Analyzed 2B+ vehicle infotainment system logs to extract insights on high-memory processes and low-impact crashes; fixes reduced crashes by 20% and memory usage by 10%.
Parsed 1B+ log lines to build a 2,000+ data-point classification dataset; applied ML classifiers with ADASYN for imbalanced data handling, achieving an F1 score of 82%.

Avoma

Machine Learning Engineer

March 2022 – July 2023 · Palo Alto, CA, USA (Remote)

Fine-tuned transformer-based embeddings for semantic similarity using bi-encoder and cross-encoder architectures (88% accuracy); deployed models via TensorFlow Serving on AWS ECS.
Implemented LLM-powered services using GPT-3.5/GPT-4 with RAG pipelines, powering 6 new product features including automated note-taking and transcript summarization.
Built an OCR-based speaker recognition system for virtual meetings on AWS Lambda, improving accuracy from 67% to 88%.
Developed a clustering algorithm over 100K+ meetings to surface top FAQs as a Django microservice on AWS EC2, increasing average CSAT from 7 to 8.
Optimized TensorFlow model inference using gRPC and AVX acceleration, achieving a 30% reduction in latency.
Set up and maintained CI/CD pipelines with TensorFlow Serving, Docker, and GitHub Actions, cutting deployment time from ~10 min to 2 min.

Helpshift

Software Engineer (Machine Learning)

July 2021 – February 2022 · Pune, Maharashtra, India

Built ML solutions and APIs for automating end-user issue resolution in customer service, including message classification and language detection.
Engineered scalable microservice inference pipelines (Django, Flask, FastAPI) supporting 1M+ requests per day.
Built Jenkins-based training and deletion pipelines for automated model lifecycle management, saving 20+ developer hours per month.
Resolved critical underperforming models due to imbalanced data, boosting classification accuracy from 70% to 90%.
Integrated logging into backend APIs for model monitoring, reducing bug resolution response time by 25%.

Segmind

Deep Learning Intern

July 2020 – October 2020 · Remote (Bangalore, India)

Contributed to CRAL, a library for abstraction of deep learning architectures for Computer Vision.
Implemented, integrated, and tested object detection models, achieving mAP scores above 0.6 on standard benchmarks.

Education

MEng, Electrical and Computer Engineering

Concentration in Applied Artificial Intelligence

University of Ottawa · September 2023 – April 2025

B.E., Electronics and Telecommunication Engineering

Pune Institute of Computer Technology

August 2017 – May 2021

CGPA: 9.45 / 10

Higher Secondary Education

Fergusson College

August 2015 – May 2017

Grade: 86.66%

Secondary Education

The Bishop's School, Pune

June 2003 – May 2015

Grade: 94.63%

Publications

Peer-reviewed research published at IEEE conferences.

DPSRGAN: Dilation Patch Super-Resolution Generative Adversarial Networks

IEEE 6th International Conference for Convergence in Technology (I2CT), 2021

Developed a novel GAN architecture to increase the resolution of images by 4x.
Achieved a MOS of 3.91/5 and PSNR of 32.24, outperforming the previously proposed SRGAN.

Publication Code

Big Data Analytics for Sustainable Cities: Pune Tree Census Data Exploratory Analysis

IEEE 11th International Conference for Computing, Communication and Networking Technologies (ICCCNT), 2020

Developed a pipeline for analysis of tree census data and introduced the Flora Biodiversity Index (FBI) to quantify tree diversity.
Drew insights to determine uniformity of tree cover and areas deficient in biodiversity.

Publication Code

Projects

A selection of things I've built.

ChargeRoute

A route optimizer for Electric Vehicles, providing charging stations while minimizing time and distance. Won first place at uOttaHack 6 (Ford Hack).

DjangoHTML/CSSJavaScript

Source Code

Cell Free Layer Detector

A deep learning image segmentation model to detect the cell free layer in blood vessels, achieving a DICE coefficient of 0.91.

TensorFlowMATLABStreamlit

Source Code

License Plate Reader

An application using YOLOv3 for plate detection and PyTesseract for OCR. Can search the database for vehicle owners and send automated notifications.

PyTorchDjangoPyTesseractSQLite

Source Code

Network Anomaly Detector

A machine learning approach to network threat detection from .pcap files, achieving 99.6% accuracy on benchmark datasets.

Scikit-LearnDjangoMatplotlib

Source Code

Survey and Rescue Drone

An autonomous drone for survey and rescue, capable of detecting beacons and maneuvering autonomously. Part of the e-Yantra Robotics Competition by IIT Bombay.

ROSOpenCV