Experience & Projects

Full‑stack, applied ML, and embedded systems — 5 years professional, 8 years building.

Shoebox Ltd.
May 2024 - Present

Core Services Developer

Shoebox Ltd.
Advanced hearing care solutions, including digital audiometers and hearing screeners that are both software-driven and cloud-integrated. Allowing professionals to efficiently conduct hearing assessments and manage patient data.

FastAPI · Python · ChromaDB · HDBSCAN · PostgreSQL · TypeScript · React · Puppeteer

  • Customer Analytics Platform:
    Architected and built a full-stack ticket analysis platform with a FastAPI/Python backend, React/TypeScript frontend, PostgreSQL persistence, and ChromaDB vector storage. Implemented VoyageAI embeddings and HDBSCAN clustering for automated categorization; designed a persistent caching layer with intelligent invalidation that reduced category-generation API costs by 90%.
  • Testing & Validation:
    Wrote pytest unit and functional tests under ANSI s3.6-2018; improved coverage for data ingestion, clustering outputs, and batch processing paths against production edge cases.
  • Cross-Stack Production Debugging:
    Reduced production calibration failures by debugging S3 bucket integrations and refining the internal Unified Calibration System APIs across C++, Python, and .NET.
  • Sound Engine Optimization:
    Diagnosed and resolved iOS sound engine regressions; collaborated with senior developers on root-cause analysis and implementation, unblocking releases.
  • Automated Testing Dashboard:
    Pitched and developed a TypeScript/Node + Puppeteer + React dashboard automating multiple testing procedures, boosting testing efficiency by ~75% as measured by reduced cycle times.
Body M3canix Inc.
Feb 2024 - Present

Full Stack Software Developer

Body M3canix Inc.
Medical-grade fitness tracker which collects precise body metrics during workouts, such as heart rate, blood oxygen, and body temperature. This device seamlessly integrates with fitness apps to provide personalized coaching and insights that help users optimize their training and reach peak performance.

Embedded C · Zephyr RTOS · BLE/GATT · PIC18F · TypeScript · React · Electron · Python · Flask · PyTorch

  • Machine Learning on Biosignals:
    Trained and productionized a PyTorch Temporal Fusion Transformer (TFT) on HR/SpO₂/temperature streams, leveraging attention mechanisms to capture long-term dependencies for early anomaly detection. Built data pipelines for feature engineering, normalization, and inference serving end-to-end.
  • Firmware & Embedded Systems Development:
    Developed C firmware for a Nordic nRF52840-based NORA-B1 (Zephyr RTOS, ubxlib) and PIC18F on a custom PCB. Partitioned tasks—PIC18F collected sensor data; NORA-B1 managed BLE (GATT) and edge processing—and implemented robust UART communication.
  • Full-Stack Platform:
    Spearheaded a TypeScript/React/Tailwind dashboard with a Python Flask backend communicating with Azure for analysis and PostgreSQL/MongoDB for storage; secure Web Bluetooth integration with live telemetry and CSV export.
  • Multi-Platform Application Delivery:
    Shipped cross-platform desktop apps (Windows, macOS, Linux) using Electron and React with MobX for state management, ensuring SOC2 compliance while leading a team of 2. Built iOS/Android mobile in React Native (TypeScript) with Zustand — all featuring secure, encrypted BLE communication with devices.
  • Technical Highlights:
    • BLE/GATT: custom 128‑bit services for metrics (SpO₂/HR, breathing, temperature), decimated raw PPG streams, command write; Web Bluetooth UI.
    • Bridge: u‑blox NORA (nRF52840, Zephyr) ↔ PIC18F26K22 over UART with ring buffers; timestamp normalization; structured diagnostics.
    • Sensors: MAX30102 (8192 ADC, 100 Hz, 400 µs) and MLX90632; LED auto‑tune to avoid saturation.
    • Algorithms: HR (peak detection, adaptive thresholds, median‑of‑RR), SpO₂ (ratio‑of‑ratios with perfusion gating), Breathing (band‑pass/envelope, 6–30 BrPM), Temperature (calibration‑aware mapping + smoothing).
    • Bootloader over BLE: control/data/status characteristics; HEX upload via web app; NORA forwards to PIC.
    • Throughput/robustness: raw PPG decimation (~50 Hz), 500 ms notification gating, contact/saturation guards; CSV export for offline analysis.
University of Bonn
May 2022 - Jul 2022

GUI Developer

University of Bonn
Built a PyQt editor for semantics-enhanced robot localization, streamlining annotation for SLAM mapping experiments. The toolchain work supported a paper in IEEE Robotics & Automation.

Python · PyQt · Robotics · SLAM