Debaditya Shome
Senior Applied AI Scientist, Toronto
I build multi-agent LLM systems that run in production, and the evals and guardrails that keep them honest. I came to it the long way: a kid who wanted his chess program to learn, then eleven first-author papers, including AAAI.
- Now
- Senior Applied AI Scientist at Homebase, since 2024
- Before
- MASc in AI at Queen’s University (Ingenuity Labs), Vector Institute, SHL Labs
- Record
- 12 publications, 490+ citations, reviewer for NeurIPS, AAAI, CVPR and ECCV
debadityaqu@gmail.com
The long version is the game on that board. Scroll, or press →, to play through it.
The game
Sixteen moves so far, each one a decision. The annotations are mine.
1. e4 A machine that learns
Before any of it
The first move is the oldest one. I grew up playing chess, and represented my school at nationals. Most of that time went into becoming the best program I could run in my own head, working through every line before I touched a piece. It taught me how to think, and I still try to see three moves ahead. I also grew up on math olympiads and on games against the computer, and I kept asking one question of it: could this program learn from my games and get better, or would it only ever follow its rules? I had not written a line of code. Everything after this is an attempt at an answer.
1… e5
2. ♞f3 Self-taught
Undergraduate, from 2018
Develops a piece, though not to the square I had planned. I chose electronics because I wanted to build robots, and was placed in telecommunications. No computer science in high school, none in the degree. So Python and machine learning were learned on the side, through online courses and internships.
2… d6
3. d4 No environment
First research project
Opens the centre. A professor wanted to try reinforcement learning for allocating network bandwidth. It was my first time reading papers, let alone writing one. There was no simulator for the problem and no repository to start from, so we spent two to three months designing the environment before any agent could run in it. Then I wrote Q-learning and deep Q-learning from scratch, years before a coding agent could have helped.
3… ♝g4
4. dxe5 Borderline
First peer review
The first exchange. We sent the paper to ICC, one of the top conferences in communications. It came back borderline, then rejected, and the reviewers were right: the experiments were thin. I designed a new reward function, reran everything against their comments, and the paper was accepted at ICCCNT. Six months from knowing nothing to a published result, and a lasting respect for what a good review can do.
4… ♝xf3
5. ♛xf3 Mitacs Globalink
Summer 2021
The queen comes out early. That paper earned a research internship in Canada on federated learning for wireless networks, and I spent the months before it teaching myself the field. On the side I saved up for 64 GB of RAM, fitted it to my own laptop, and ran reinforcement learning in the CARLA driving simulator and in bots for the video games I played. It was slow. It was also how I found computer vision.
5… dxe5
6. ♝c4 FedAffect
ICCV Workshop, 2021
The bishop takes its best diagonal. The first idea that was entirely mine: use federated learning to recognise facial expressions from very few labelled examples. I took it from nothing to an accepted ICCV workshop paper in a few months, alongside the internship. I never got to continue it, and it is still one of my favourite pieces of work.
6… ♞f6
7. ♛b3 SHL Labs
October 2021 to April 2022
A double attack: a paper and a tool from the same idea. At SHL Labs I spent months on in-context learning, before anyone called it prompting, trying to beat the domain-specific BERT encoders the team relied on. The result was EXnet, a model for in-context text classification that outscored GPT-Neo-1.3B on F1 at roughly one eighty-sixth of the size, and it went into an active-learning tool for labelling data. It was a very good time, and it is where I learned how language models work, one era before LLMs.
7… ♛e7
8. ♞c3 Queen’s and AAAI
September 2022 to February 2024
Quiet development that wins the game later. I could only take a graduate offer that paid for itself, and Queen’s did: a fully funded MASc in AI at Ingenuity Labs, on a Vector Scholarship, a Mitacs Globalink Graduate Fellowship and the university’s own award. My first problem there, generating one physiological signal from another, took a year. The result is a region-disentangled diffusion model that reconstructs an ECG from the optical pulse a smartwatch already records, cutting error by 86 percent against diffusion baselines. It went to AAAI, and it is my best research.
8… c6
9. ♝g5! The salvage
ICASSP and ICMI, 2024
The pin. Next I tried emotional speech synthesis with diffusion models. You could hear the emotion, but without the compute to scale it the voices stayed robotic, and the project failed. One component held up on its own: distilling prosodic and linguistic cues into a single encoder for emotion recognition. That became the ICASSP paper. The thesis closed with mood classification from sleep and wearable data, at ICMI, and a 4.15 out of 4.3.
9… b5
10. ♞xb5! Vector Institute
Autumn 2023 to spring 2024
The sacrifice. An internship at the Vector Institute put language models in my hands: a course recommender built on dense retrieval and re-ranking, with the bi-encoder fine-tuned on labels written by LLMs. Back in the lab, my last months went to multimodal models, trying to build a better LLaVA. The plan had been a PhD. I gave it up for the initiative, because I wanted to build with these models and not only publish about them.
10… cxb5
11. ♝xb5+ Homebase
July 2024
Check. Homebase was starting an applied AI team, I liked where they were going, and I joined as its first applied AI hire. The first build was the Hiring Assistant: I was the sole applied scientist on a seven-agent, human-assisted screening workflow, along with its screening evaluations, guardrails and red-teaming.
11… ♞bd7
12. O-O-O! Evals and guardrails
2025
King safety before the attack. Shared evaluation infrastructure for every agent product, with LLM judges calibrated against people and failures attributed to the agent that caused them. Then a fine-tuned open-source guardrail model, deployed in front of production agents. Castling also puts a rook on the open file.
12… ♜d8
13. ♜xd7 Image2Schedule
2025
Gives up the exchange to keep the pressure. A vision-language model fine-tuned end to end to read a photograph of a paper schedule, with a task-specific metric and an automated evaluation loop to iterate against.
13… ♜xd7
14. ♜d1 Homebase Assistant
2025 to 2026
The last piece joins. A constrained-autonomy multi-agent harness: query planning, parallel tool execution, isolated interpretation of results, answers that cite their sources. When the agent acts, the owner’s approval is bound to the exact tool call. Its tool-design standard was adopted across engineering.
14… ♛e6
15. ♝xd7+ Senior
September 2025
Recaptures with check. Promoted to Senior Applied AI Scientist, fourteen months after joining.
15… ♞xd7
16. ♛b8+!! Pioneer Award
2026
The queen sacrifice the whole game was played for. Homebase’s highest internal award, given to six people across the company.
16… ♞xb8