Hi, I'm Sai .
I work on generative photo and video models at Canva 1 Senior Applied Scientist , Video Studio. Before that, Photo Effects, where I worked on Background Generator (1M+ monthly users)., and outside of that I like taking things apart to see why they work: why anything associative can be split in half , how every trainer in Pokémon Red would rank against each other, what's actually sitting inside Yakuza 0's data tables .
This site is where I write it down. Some of it is maths, some is ML systems, and a lot of it is advice for students trying to get into Australian tech 2 First Class Honours , Applied Data Science at Monash; thesis on bias in diffusion models. Once president of the Monash Association of Coding., written from not that long ago.
Recently
- A GPU-performance reading map for making deep learning faster: hardware hierarchy, rooflines, kernels, profiling, distributed training, serving, quantization, and monokernels.
- Exploiting associativity on binary operators to speed up operations
- Decoding Yakuza 0's data tables
- Framing recursion and dynamic programming as mathematical induction, with worked examples from linked lists and binary trees.
- A summary of some software I've been working on, and other updates.
Currently: probably napping.
Trainer card
Trainer Card No. 025
Sai Kumar M.K.
Senior Applied Scientist · Canva
MLHPCResearchMathsReverse eng
- Posts
- 27
- Dex
- 15
- Badges
- 4
- Since
- 2021
Badges Select to explore
Project dex · highlights
No. 001
monash-handbook-plus
Unit search, cost calculator, areas of study browser, an interactive prerequisite graph, a degree planner with auto-scheduling, a pathway finder and shareable study plans.
No. 002
saikumarmk.com
This site: writing, the Yggdrasil skill tree, the project dex and a few 3D visualisations.
No. 003
mini-melbourne-3d
A new Mini Melbourne: buses, trams and V/Line in 3D.
No. 004
pokered-trainer-tournament
A recreation of the Pokémon Red Elo tournament on the pkmn engine.
Experience
Currently Senior Applied Scientist at Canva .
- Built an agent evaluation platform with automated scoring, execution traces, and a results viewer, enabling research and engineering teams to inspect changes, compare models, and evaluate release quality independently.
- Evaluated candidate models for agent and generative-media tasks, comparing output quality and inference cost; used automated assessments for frequent feedback and human reviews ahead of releases.
- Modernized transcription serving with vLLM and forced alignment, cutting processing time for an hour of video from approximately five minutes to one while improving recognition accuracy and word timing.
- Reduced LLM inference cost by 46% on a conversation benchmark through prompt and tool-definition caching; shipped the optimization to production.
- Removed a document-validation bottleneck affecting every editing operation, reducing validation latency from seconds to milliseconds by eliminating repeated Python stack inspection.
- Built a headless Linux version of a motion-graphics engine and ported its build from CMake to Bazel, making it available for integration with an export service.
- Served as a technical point of contact between research and product teams, reviewing model integrations, unblocking serving changes, and helping engineers run evaluations independently.
- Integrated support for multiple LLM providers into the agent orchestration stack, enabling model comparisons and experimentation.
- Developed detail-preserving methods for Background Generation, a feature that reached over 1M monthly active users; named inventor on the resulting filed patent application.
- Halved end-to-end outpainting latency for Magic Expand by deploying distilled SDXL and keeping image processing on the GPU, while improving visual quality in human evaluations.
- Adapted a SLURM/
torchruntraining stack to Ray/Anyscale, replacing local-data assumptions with an S3-streaming loader and validating distributed training with platform and research teams.
- Developed and evaluated bias-mitigation methods for Stable Diffusion in Canva's text-to-image product, focusing on representation across gender and ethnicity.
- Presented internship research as part of Generation Gap: Addressing Bias in Generative AI, SIGGRAPH Asia 2023.
2022 Data Science Intern DiGiCOR Led a team of four to build a sales dashboard using Streamlit and NetSuite's REST API.
2022 Winter Research Assistant Monash Built a genetic-programming and simulation framework in Julia to investigate a game-theoretic problem in scientific publishing.
2021 – 22 Summer Research Assistant Monash Analyzed commit data from more than 30,000 deep-learning repositories using Python and the GitHub API to study the evolution of ML libraries for AutoML.
- Oversaw growth to over 1,100 members and ran a technical careers evening with over 100 attendees.
- Ran four Python workshops covering FastAPI, discord.py, and web development, reaching over 200 attendees.
- First Class Honours (final honours grade: 90); Dean's List for academic excellence in 2021, 2022, and 2023.
- Thesis: Bias Modelling and Mitigation in Diffusion Models.
The one-page version: printable CV .