If infinitely loading in center, refresh page
01 — about

I'm a Mechatronics Engineering student at the University of Waterloo interested in robotics, ML, learned motion, and the software that makes physical systems behave intelligently.

At this stage, I look to optimize for depth over breadth. The intersection of hardware and learned behavior is where I want to build, and I'm spending this window figuring out exactly how to get there.

Currently studying full-time while building foundations in systems programming and ML, with my sights set on robot learning and autonomous systems.

  • 01DisciplinesSoftware · Machine Learning · Robotics
  • 02StackC++ · Python · Pytorch · ROS2
  • 03Looking forSWE / Robotics ML/AI internships, Fall '26
  • 04Lovesautonomous systems · the art of a good meal · carnatic music (performing)
02 — selected work
JAN 2026 - APR 2026

AI Software Engineer Intern

CGI Inc.

Trained a priority classifier on 280 ITSM tickets, tuning for weighted F1 and Critical ticket recall under 7% class imbalance Built duplicate detection via sentence embeddings and cosine similarity, calibrating threshold on 90 hand-labeled pairs to 0.86 precision Clustered 240 failure embeddings with DBSCAN, surfacing 3 actionable root causes for Foresters Financial Closed a data flywheel that improved ticket classification from 76% to 84% across retraining cycles

  • SKLEARN
  • DBSCAN
  • PYTHON
  • AZURE/AWS
01
JUL 2025 - AUG 2025

Cloud Engineering Intern

Avanade Canada

Architected a cloud-native AI accessibility platform across 8+ Azure services with a 6-person team Built an intelligent support chatbot covering 15+ disability categories with sentiment analysis and escalation routing Cut provisioning time by 30% with infrastructure as code and automated CI/CD pipelines

  • AZURE
  • DOCKER
  • PYTHON
  • GITHUB ACTIONS
02
OCT 2024 - MAR 2025

Controls Engineer

UWaterloo

Designed a mechanical airbrake system in SOLIDWORKS, validated via MATLAB flight simulation Built a custom PCB airbrake controller in KiCAD with signal processing to filter noisy telemetry during active flight

  • SOLIDWORKS
  • MATLAB
  • KICAD
  • C++
03
03 — projects
  1. 001
    2026

    Smart Pressure Sensing Insoles HACKPRINCETON SPRING ’26 FINALIST

    Clinical pressure mats cost $15,000. We built a directionally accurate version for $77 because our parents are diabetic and can't feel the pressure that destroys feet. Hand-soldered 10 FSR sensors, analog multiplexer, haptic motors, and transistor drivers onto an ESP32, streaming pressure data at 15Hz over BLE Derived clinical metrics — arch index, pronation, heel centering — computed locally on-device, interpreted by a Gemini-powered AI podiatrist Haptic alerts fire when sustained pressure exceeds threshold, running in background via CoreBluetooth

    Github ↗
  2. 002
    APR 2026

    Natural Language  PCB Designer

    PCB design tools cost thousands and assume you already know electronics. Boardsmith doesn't. Describe a circuit in natural language or drop a breadboard photo and Gemini parses intent into a structured circuit, then a custom Lee-algorithm autorouter lays copper on a real board Streams 8 pipeline stages live via SSE: schematic, firmware, BOM, placement, routing, 3D render, Gerber bundle, all manufacturable by JLCPCB Generates pin-aware Arduino firmware, SPICE simulation links, and a real-time cost calculator from 1 to 500 boards

    Github ↗
  3. 003
    APR 2026

    Visual Inventory Catologer

    Shopify merchants spend days manually listing physical inventory online. VisualShelf cuts it to one click. Indexes a merchant's entire Shopify catalog as CLIP embeddings in Qdrant — photograph a product from the stockroom and cosine similarity search finds the match instantly For unmatched products, GPT-4o generates a title, description, and tags from the image and publishes the listing directly via Shopify Admin GraphQL Evaluated against a held-out test set with augmented images, precision@5 above 0.80 and catalog coverage above 0.75

    Github ↗
  4. 004
    FEB 2026

    PR Risk Intelligence Platform

    Engineering teams find out a PR is stuck after it's already been sitting for three days. This flags it before it stalls. 3 decoupled Python microservices on Azure Container Apps ingest 500+ daily GitHub webhook events, deduplicate via Redis, and aggregate PR lifecycle metrics in PostgreSQL Gradient boosting classifier trained on 10,000+ GitHub events predicts merge time risk at 82% accuracy, with features including review assignment lag, code churn, and author velocity At-risk PRs surface in a React TypeScript dashboard before bottlenecks happen, not after

    Github ↗
04 — contact

Let's make something —

s2yeshwa@uwaterloo.ca

Tweaks