Software Engineer, AI and Automation Practice, IBM · Toronto, ON
I build AI infrastructure and the tools around it: self-hosted model serving, browser agents, and data platforms that reach production.
Portfolio · CV (PDF) · LinkedIn · Email
flowchart LR
A[Model serving<br/>vLLM, FP8, MTP] --> C[Applied AI systems]
B[Browser agents<br/>CDP, MCP, Manifest V3] --> C
C --> D[Evaluation<br/>QLoRA, GRPO, 1,240 tasks]
D --> A
Model serving: vLLM, SageMaker, FP8 weight compression, tensor parallelism, speculative decoding (MTP), and KV cache tuning.
Browser agents: raw Chrome DevTools Protocol, MCP servers, and Chrome Manifest V3 extensions.
Applied AI: QLoRA and GRPO fine-tuning, a six-agent text-to-SQL platform, and evaluation of structured outputs.
- Data platforms: BigQuery, Snowflake, ETL, Pandas, Terraform, and CI/CD.
- Languages: Python, SQL, TypeScript, JavaScript, R, and Rust.
IBM, Toronto · May 2025 to present
- Deployed Gemma 4 31B as a self-hosted SageMaker endpoint with vLLM 0.32.0, FP8 weight compression and tensor parallelism. The deployment kept regulated data in the VPC and removed the managed-model dependency.
- Benchmarked a 31B vLLM deployment on eight A100 GPUs. Throughput reached 10,410.60 tokens per second at N=512 with zero errors at N=1024, and cost fell from $132.56 to $0.77 per million tokens, a 172x reduction.
- Ran 144/144 successful A/B runs of multi-token prediction across A10, A100-80GB, H100 and B200 GPUs. A10 was the only GPU with a durable MTP win, at 1.38x to 2.06x.
- Delivered a six-agent text-to-SQL platform with 93% accuracy across five production databases in eight weeks, enabling non-SQL staff to query operations data.
- Built a five-agent cross-cloud A2A platform with Entra ID JWT propagation, AWS Lambda SigV4, Bicep and Arize Phoenix. It has 124 unit tests and 16 RBAC roles, and it cut API calls 73%.
- Optimized H200 kernels for Muse Spark: 1.53x dense projection, 1.96x MoE W4A16, 1.90x attention decode and 1.77x decoder.
Toronto General Hospital, UHN · Sep 2024 to May 2026
Supervised by Dr. Mamatha Bhat and Dr. Joseph S. Kim. Built a liver-failure survival model that reached 88.7% accuracy on research data from 1,260 patients across 32 US centers, and multimodal transplant research with early fusion of clinical text and histopathology features.
University Health Network (UHN), KITE Research Institute · May 2024 to August 2024
Human activity recognition with limited input data, supervised by Dr. Jose Zariffa. Research models recognized human actions with over 68% accuracy across different test settings.
Sanofi, Toronto · January 2024 to August 2025
Built a two-click Snowflake and Streamlit workflow that moved manual reporting from days to seconds, and led a cross-functional team that introduced sprint planning and CI/CD automation, supporting a 40% team productivity gain.
Leslie Dan Faculty of Pharmacy (LDFP), University of Toronto · May 2025 to August 2025
Supervised by Professor Mina Tadrous. Built a paper-screening agent with OpenRouter, Gemini and OpenAI models, and a pharmacist-patient platform with NodeJS, ExpressJS, NextJS, React and TypeScript.
Acceleration Consortium · May 2025 to August 2025
Research tools extracted structured data from more than 10,000 document pages with 99.98% accuracy, producing more than 10,000 data rows with 22 fields each.
University of Toronto · May 2023 to December 2023
A financial dashboard made three years of performance easier to review and compare across time.
University of Toronto · September 2023 to April 2026
Taught calculus, data science and probability for 1,029+ hours, led tutorials for 30+ students, ran review seminars for 400+ students, and graded assessments for 650+ students.
University of Toronto
A university vision study reached 83% overall accuracy when evaluating object recognition and grasping together. A separate health project studied gendered health through tracking and body sensors.
Kimia Farma · September 2022 to December 2022
Built a branch sales database of 140,000+ entries and a Tableau sales dashboard, reducing data-processing time by 33%. The dashboard saved $95,000 in the annual budget.
Mulia Health Dental Care Group · March 2022 to July 2022
Customer-data analysis across 14 branches increased operational efficiency by 15%, using data from more than 10,000 customers. Recommendations informed initiatives that supported a 25% increase in company profitability.
- OpenBrowser: a Python framework that drives Chrome over raw CDP and runs LLM-written Python in a persistent namespace. It ships a CLI, a single-tool MCP server, a persistent daemon, a FastAPI and Next.js web app, and 14 LLM provider adapters. CLI and MCP comparison
- Context Atlas (Jev + Laya): a Manifest V3 Chrome extension that finds the sentence that answers a question on the current page and highlights it, with local search by default. Read the article
- Multi-Token Prediction benchmark: 144/144 runs succeeded for Gemma 4 E2B-it at batch size 1 across A10, A100-80GB, H100 and B200 GPUs. The study separates engine gains from speculative-decoding gains.
- Browser-in-the-Loop: Reinforcement Fine-Tuning LLM Agents for Web Form Filling (STAD68 final paper, 2026): QLoRA and online GRPO raised the validation-task score from 0.614 to 0.670 (+9.1%) on 1,240 browser form tasks. These are benchmark results, not measured production savings.
- Concentrate or Collapse: When Reinforcement Learning Meets Diffusion Language Models for Web Planning (STAD80 course paper, 2026): compared Qwen3-8B, ReFusion-8B and FS-DFM-1.3B on 1,240 FormFactory tasks. FS-DFM scored 0/124 greedy despite 76.6% to 87.9% stochastic parser validity, which shows that parse-valid plans can still fail in a browser.
- University of Toronto: Honours Bachelor of Science with High Distinction, June 2026. Data Science and Machine Learning Specialist, Computer Science major, Economics minor. CGPA 3.87/4.00.
- Georgia Institute of Technology: Master of Science in Computer Science, Artificial Intelligence Specialization, 2026 to present.
- Awards: Dean's List (2023 to 2025), T-CAIREM Research Award, and Advanced Indonesian Scholarship.





