Developer

All 19 entries, newest first. Pick a focus or a skill to move the matching work to the top. Open a row for outcomes, screenshots and links.

Select

Everything, newest first

Roles, projects and hackathons.

01Web · ML / AIShoebox Ltd.UCS ownerEmployerPython · FastAPI · .NET · TypeScript · React · AWS S3 · Claude Code · MCP · pytest May 2024 to nowShoebox site

I work on Shoebox's hearing-test devices and the cloud services behind them. I own the Unified Calibration System (UCS), Shoebox's internal calibration software, a Python service built on the .NET APx API and S3. I took it over from the previous lead in July 2026. I also run iOS release sanity testing, write IQ, OQ and PQ validation protocols, and build internal tools such as a two-agent PR review skill.

  • Own the Unified Calibration System (UCS) since July 2026. I make the updates, run verification and validation, and release to production, most recently UCS 2.6 with the APx500 v10 upgrade and the comparison calibration that approved it
  • Caught a UCS atmospheric-pressure limit that didn't match ANSI S3.6-2018, filed the fix and audited 7,600+ calibration records from two years to size the impact
  • Showed CI coverage reported as 98% was really 53% (the login and upload scripts had no tests), then planned the fix
  • Cut mic-test reporting from a full 8-hour day per sanity cycle to 30 minutes with a Python proof of concept
  • Built a two-agent Claude Code PR review skill. An evaluator drafts the review from the diff and the team's standards, pulled from Confluence over MCP. A vocalizer rewrites it in the reviewer's voice, and after each review the skill compares its draft with what the reviewer actually posted and feeds the difference back
  • Built a ticket analytics platform on FastAPI and React with VoyageAI embeddings, ChromaDB and HDBSCAN clustering. Persistent caching cut category-generation API costs by 90%
  • Ran sanity testing on 12 iOS and iPadOS releases and led 7 of them
SYSTEM SKETCH · SHOEBOXCALIBRATION · UCS, OWNER SINCE JUL 2026TICKET ANALYTICSPR REVIEW · CLAUDE CODE SKILLAPX500 V10.NET APx APIUCSPython · ANSI S3.6S3calibration recordsVOYAGEAIembeddingsCHROMADBvectorsFASTAPIHDBSCAN · cacheREACTdashboardCONFLUENCEstandards · MCPEVALUATORreads diff, draftsVOCALIZERreviewer's voiceREVIEWERedits, postsEDITS FEED BACKmy codehardware / service

Swipe to see the whole sketch →

02Firmware · Web · ML / AIBody M3canixPart-time · Atalanta wearableEmployerC · Zephyr RTOS · BLE/GATT · PIC18F · React · React Native · Electron · Flask · PyTorch Feb 2024 to Aug 2026Atalanta site

Atalanta is a chest-strap fitness tracker that measures heart rate, SpO₂, temperature and breathing rate. I led engineering part-time alongside Shoebox, wrote the firmware for both chips myself, and led a team of 2 on the web app, the backend and the desktop client.

  • I wrote the firmware for both chips myself: the BLE bridge on a Nordic nRF52840-based NORA-B1 (Zephyr RTOS) and the sensor firmware on a PIC18F on a custom PCB, linked over UART
  • Wrote the on-device signal processing: dual-LED PPG on a MAX30102 at 100 Hz, heart rate by adaptive-threshold peak detection, SpO₂ by ratio of ratios with perfusion gating, and breathing rate from a band-pass envelope
  • Kept the BLE telemetry stream steady under poor signal with ring buffers and timestamp normalization, and held throughput with decimation and 500 ms notification gating
  • Led a team of 2 on the web app, the Flask backend and the Electron desktop client. I built the React Native (Expo) mobile app myself. Every client talks to the device over encrypted BLE and to the same backend
  • Trained and productionized a PyTorch Temporal Fusion Transformer on heart rate, SpO₂ and temperature data for early anomaly detection, with the feature, normalization and inference-serving pipeline
  • Set up CI/CD on GitHub Actions with Docker, and Azure for deployment and data storage
SYSTEM SKETCH · BODY M3CANIXDEVICE · CUSTOM PCBCLIENTSBACKENDMAX30102+ MLX90632 sensorsPIC18Fsensor firmware · CNORA-B1Zephyr · BLE bridgeWEB APPReact · Web BluetoothMOBILEReact Native · ExpoDESKTOPElectron, led 2 devsFLASK APIPythonPOSTGRES · MONGOon AzureTFT MODELPyTorch · anomaliesI²CUARTBLE GATT · ENCRYPTEDmy codehardware / serviceI WROTE THE FIRMWARE · LED 2 ON WEB

Swipe to see the whole sketch →

Scopes home screen: Ottawa city status and daily summariesScopes alerts screen with Environment Canada weather warnings
03Web · ML / AIScopes, city status & alertsProjectMy projectTypeScript · Node · Firestore · Vercel · OpenAI · React Native 2024GitHub

One screen for Ottawa's city service status, weather, traffic, police updates and local news. I built all of it: the scrapers, the API, the LLM summaries and the mobile app.

  • TypeScript scrapers (RSS, REST and Puppeteer) run on schedules in GitHub Actions, every 30 minutes to every 6 hours, and write to Firestore through one repo layer that keeps dev and production data apart
  • Built serverless Node and TypeScript REST APIs on Vercel with strict schemas and date-based indexing
  • Daily two-bullet summaries from the OpenAI API, plus headline enrichment and road-based filtering, cut manual triage to near zero
  • Built the React Native (Expo) app with typed Apisauce clients, FlashList lists, pull-to-refresh, deep links, and 7 languages with right-to-left support
  • Designed the app UI in Figma
Scopes home screen: Ottawa city status and daily summariesScopes alerts screen with Environment Canada weather warnings
04WebSpotify Playlist ManagerProjectMy projectTypeScript · React · Firebase · OAuth Feb to May 2024Live app

A web app for editing very large Spotify playlists (10,000+ tracks): merging, removing duplicates and bulk edits.

  • Built the front end in Vite, React and TypeScript with Tailwind and Radix UI. Virtualized lists render only the visible rows of a 10,000+ track playlist
  • Used Spotify OAuth with chunked writes to stay under rate limits, and the Web Playback SDK for track previews
  • Wrote a Firebase Cloud Functions (Express) backend that keeps client credentials out of the browser and generates preview URLs
Spotify Playlist Manager logo
05Robotics · ML / AICarleton Planetary Robotics TeamRobotics perception & manipulationMemberROS2 · Jetson AGX Orin · YOLOv7 · C++ · Python · Gazebo Sep 2023 to Apr 2024CPRT site

The Carleton Planetary Robotics Team builds a Mars rover. I worked on the arm's perception and manipulation, running ROS2 Humble on a Jetson AGX Orin with a ZED 2 stereo camera.

  • Integrated the ZED 2 with ROS2 Humble on the Jetson: camera drivers, depth sensing and point clouds
  • Built Python and C++ ROS2 nodes that chain YOLOv7 detection, depth-based 3D localization, inverse kinematics and motor actuation on a 6-DOF arm
  • Wrote the IK solver with homogeneous transforms and validated it in Gazebo Ignition and in driver-level hardware tests
  • Wrote the arm's motor driver nodes, contributed a ROS wrapper for the ZED SDK and supported the SLAM work
SYSTEM SKETCH · CPRT ROVER ARMPERCEPTION · ROS2 HUMBLE ON JETSON AGX ORINMANIPULATIONTOOLING · VALIDATIONZED 2stereo · depthYOLOV7 NODEobject detection3D LOCALIZEdepth → grasp target6-DOF ARMhardwareMOTOR DRIVERSROS2 nodesIK SOLVERhomogeneous transformsZED WRAPPERROS · SLAM supportHW TESTSdriver levelGAZEBOIK in simulationmy codehardware / serviceCORE MEMBER · 2023 TO 2024

Swipe to see the whole sketch →

Title slide of the clean-cooking campaign research deck
06ML / AIUNEP x TKS campaign researchProjectCollaborationPython · BeautifulSoup · Data analysis Mar to Aug 2023Campaign deck

Through The Knowledge Society, I led a 4-person team that designed a global YouTube campaign for UNEP to fund clean cooking solutions for rural households in India.

  • Mined YouTube, TikTok and Instagram data with Python and BeautifulSoup to guide targeting and campaign design
  • Interviewed 10+ experts on the challenges and openings for clean cooking in rural India
  • Pitched the campaign strategy to UNEP project managers
Title slide of the clean-cooking campaign research deck
Mission Control
07RoboticsMission Control telemetry processorProjectC++ · UDP · POSIX 2023GitHub

A C++ telemetry service for a lunar-lander challenge. It reads binary UDP packets from the lander and detects the landing event.

  • Used non-blocking POSIX sockets for high-throughput UDP
  • Strict binary message packing, endianness handling and sensor scaling feed the landing-event logic
Face detection boxes drawn on two press photos
08ML / AIInvest Ottawa (client)ImageClusterBlur photo triagePython · PyTorch · MTCNN · fastdup Dec 2022 to Mar 2023GitHub

A Python CLI that helps photographers at Invest Ottawa sort a shoot. It finds faces, groups near-duplicates and flags blurry frames, and made client review about 60% faster.

  • Combined MTCNN face detection, fastdup near-duplicate clustering and variance-of-Laplacian blur scoring in one modular, class-based pipeline
  • Added GPU batch processing with configurable thresholds
  • Added RAW to JPEG conversion and automatic file organization, with error handling and logging
Face detection boxes drawn on two press photos
A ship painting re-rendered in several styles
09ML / AIStyle transferProjectPyTorch · VGG19 · CycleGAN Mar to Dec 2022GitHubPaper

Neural style transfer in PyTorch, following Gatys et al. with all 16 VGG19 layers and the shift-activation improvement from Novak and Nikulin. I also trained a CycleGAN for unpaired image translation.

  • Wrote a fast style transfer on VGG19 feature stacks and Gram matrices that optimizes the image with Adam and saves periodic outputs
  • Trained a CycleGAN on horses and zebras, with two generators, two discriminators, adversarial and cycle-consistency losses, mixed precision and periodic checkpoints
A ship painting re-rendered in several styles
PyQt semantic map editor with a labeled floor plan
10RoboticsUniversity of BonnGUI Developer (Robotics)Python · PyQt · SLAM May to Jul 2022IEEE paperGitHub

At the University of Bonn, I led development of a cross-platform PyQt editor that researchers use to build abstract semantic maps for robot localization.

  • Refactored and sped up the legacy code to add new features, working closely with the researchers
  • Extended and stabilized the toolchain used in a SLAM study, and wrote the documentation and software support for its IEEE Robotics and Automation Letters paper
PyQt semantic map editor with a labeled floor plan
CanCare logo
11WebCanCareHackathonFull-stack · DeSo Mar 2022Devpost

An app that tracks a cancer patient from diagnosis through treatment. Built on the DeSo blockchain with a team of 5 for HealthHacks by PennApps x Wharton, where it won Popular Choice and Best Use of DeSo.

CanCare logo
Grid of GAN-generated spider images
12ML / AISpiderGANProjectPyTorch · GAN Jan to Mar 2022GitHub

DCGAN and WGAN in PyTorch, trained to generate spider images.

  • Built the WGAN critic with weight clipping and several critic steps per generator step
  • Wrote a custom SpiderDataset with ImageFolder support and a normalized 64×64 pipeline
  • Trained with RMSprop and Adam, logged to TensorBoard, with checkpoints and image grids saved for review
Grid of GAN-generated spider images
A frame of the Snake game the agent plays
13ML / AISnakeGameAIProjectPyTorch · DQN Feb to Mar 2022GitHub

A deep Q-network agent that learns to play Snake and reaches an average reward of up to 30.

  • Built the agent with a Nature-CNN encoder, a 1M-transition replay buffer, a target network and epsilon-greedy exploration
  • Vectorized training across environments with SubprocVecEnv and frame stacking
  • Added msgpack model serialization, and TensorBoard logs for reward and episode length
A frame of the Snake game the agent plays
The sandbox UI with a prompt and a generated answer
14Web · ML / AIInstructGPT3 SandboxProjectReact · OpenAI API · Python Feb 2022GitHub

A web app for answering discussion questions with OpenAI's GPT-3 Davinci model, built in early 2022 before ChatGPT existed.

The sandbox UI with a prompt and a generated answer
NeuroVoice pitch graphic of a head wearing an EEG cap
15ML / AINeuroVoiceHackathonBCI · EEG · Transformer Feb 2022Pitch deck

A speaking-device concept built with a team of 5. A custom EEG cap sends brain signals to an autoregressive transformer that turns them into speech. I built the early prototype in a 24-hour hackathon and researched the ML models.

NeuroVoice pitch graphic of a head wearing an EEG cap
CNN cells
16ML / AICNN cell classifierProjectKeras · CNN 2022GitHub

A Keras CNN that tells parasitized cells from uninfected ones. I curated the image dataset, trained the model, exported it as CNN_cells.h5 and documented it in a notebook.

Gestly title graphic with its sign-language mascot
17WebGestlyHackathonNLP · Accessibility Dec 2021Pitch deck

A concept for a live text-to-sign-language translator that plugs into video platforms, corporate training videos and games.

Gestly title graphic with its sign-language mascot
Render of the Project Nural headband
18ML / AIProject NuralHackathonBCI · Machine learning Nov 2021Pitch video

A non-invasive BCI headband concept for treatment-resistant depression, using temporal interference stimulation. I led the machine-learning work in a team of 4, and we pitched it to investors.

Render of the Project Nural headband
Render of the CIBASTRA modular food scaffolds
19SystemsNASA Space Apps, Global NomineeHackathonRecognitionSystems design Oct 2021Project page

CIBASTRA, a modular food system for long space missions. Single-use scaffolds with a water-attracting coating hold microgreen seeds, and each module yields produce every 7 to 9 days. Named a Global Nominee.

Render of the CIBASTRA modular food scaffolds

Email me.

I'm open to full-time and contract roles in firmware, web, ML and robotics. I'm in Ottawa and can work remotely or relocate.