Portfolio / 2026
ML · Software · Product
Putting ideas into motion,from intuition to execution.
Systems Design Engineering @ UWaterloo
01 / Experience route
Route stop
Built LLM-powered learning insight workflows around educational performance data, processing pipelines, and backend reliability.
Workflows supporting 150+ educators
Route stop
Built and deployed a RAG-based grant proposal system that turns planning documents and public data into policy-aligned funding recommendations. Combined hybrid search with source validation and explainability to flag unsupported or low-confidence outputs.
Used across 25+ municipalities · Led 2 engineering sprints
Route stop
Developing agent workflows that combine LLM reasoning, tool execution, and external data to automate multi-step business processes. Building Python data and ML pipelines on Databricks and AWS, with modular orchestration for traceable, testable client implementations.
02 / Selected work
Focused systems across student workflows, civic planning, health navigation, campus coordination, AI operations, and data infrastructure.
01 / 08
Probabilistic interpretation can help, but human-approved actions must control every write.
Turns lectures, deadlines, and project meetings into structured, human-approved actions across Notion, Google Calendar, and GitHub.
One student network for the everyday logistics of campus life.
A trusted campus network bringing rides, marketplace listings, study groups, and messaging into one place.
50+ early users
Turning scattered city signals into explainable planning decisions.
A planning prototype for comparing growth, mobility, infrastructure, policy, and strategic readiness across sample urban areas.
Making clinical-trial discovery easier to navigate and understand.
Decision-support for patients and care teams discovering, comparing, and understanding relevant clinical trials.
$1,000 innovation award
Giving AI workflows somewhere reliable to run.
A system for queuing, executing, validating, retrying, monitoring, and inspecting asynchronous LLM-powered work.
Confidence-gated ETL for messy merchant catalog data.
Deterministic mapping, human review, and one publish policy turn inconsistent merchant CSV feeds into canonical product data.
Owning the full pipeline, from raw webcam capture to a trained detector, instead of wrapping an off-the-shelf model.
A dual-head CNN trained end-to-end on a custom-collected, custom-annotated dataset to detect and localize faces live from a webcam feed.
Turning hours of unlabeled field recordings into a per-recording call count with a compact CNN.
An end-to-end audio classification pipeline that detects and counts Capuchinbird calls in raw rainforest field recordings.
252 calls detected across 61 of 100 recordings
03 / Capabilities
Tools matter most in the context of what they make possible.
Ship useful interfaces and full-stack products.