Jack Shkifati
A soft glowing hologram of Jack, beamed up from a small projector chip on his portfolio circuit board

Portfolio board · Rev 3.0 · New York

Jack
Shkifati

Embedded Systems & Machine Learning Engineer

AI voice

Senior AI/ML & Software Engineer at Live Lion Security. Before that, R&D at PepsiCo: firmware, edge vision and production ML in four countries.

Power on See the chips

Every chip on this board is one of my projects. Scroll to follow the signal, and tap a chip to play its demo.

U0 · JS-MCUAbout

The microcontroller behind it all

Portrait of Jack Shkifati
Fig. 01 · Jack ShkifatiNew York, NY

I’m an embedded systems and machine learning engineer based in New York. When I’m not poring over data structures or tinkering with microcontrollers, you’ll find me exploring the city’s food scene, diving into the latest physics book, or chasing the perfect espresso.

I have 3+ years of shipping production firmware and ML. At PepsiCo I led software and computer vision for smart vending in France, Australia, India and Poland (100+ units per market), invented a patent-filed Dynamic Product Label device, and took models from PyTorch to the edge with ONNX and OpenVINO. Today I’m a Senior AI/ML & Software Engineer at Live Lion Security.

Read more

My work spans bare-metal MCUs (ATmega, STM32, ESP32) through embedded Linux, low-latency C++ and Python at the edge, concurrent backends and IoT telemetry over MQTT and AWS. I’m a C/C++ specialist who is equally at home in a PyTorch notebook or a Postgres schema.

I’ve always had a deep-rooted curiosity about how things work, not just in software and electronics but in the fundamental laws that govern our universe. Physics has always been close to my heart, and grounding my work in those principles gives me an unusual appreciation for the technology I build.

Y1 · CRYSTALExperience

The clock: where I’ve worked

  1. Senior AI/ML & Software Engineer

    Live Lion Security · Remote video security monitoring

    Aug 2026 — PresentVideo security

    Live
    Client platform
    Pilot
    Rolling out to sites
    Scale
    Multi-site by design
    • Live: client platform
    • In progress: Salesforce
    • In progress: monitoring portal
    • Next: agentic vision AI
    • Built and launched the company’s client platform, live at clients.livelionsecurity.com: a custom web app that replaced the per-site WhatsApp groups that were overloading the monitoring team.
    • Alarm and event intake from customer camera systems (Hikvision NVRs) through a device gateway, escalation workflows with push notifications, and per-site configurable camera setup.
    • Stack: Supabase (Postgres, Edge Functions, row-level security, pgTAP tests), a Cloudflare-hosted frontend, an AWS-hosted device gateway, GitHub Actions CI and nightly backups.
    • Now piloting with client sites, built to scale.
    2 more at Live Lion Security
    • In progress: building the company’s Salesforce infrastructure and its operator monitoring portal.
    • Next on the roadmap: agentic vision AI models for monitoring.

  2. R&D Software & AI Engineer

    PepsiCo · R&D

    Jul 2023 — Jan 2026Valhalla, NY

    100+
    Units per market
    20%
    Hardware BOM cut
    99.2%
    Subsystem uptime
    ~88% → 99%
    Classification accuracy
    • FRFrance
    • AUAustralia
    • INIndia
    • PLPoland
    • Patent filed: Dynamic Product Label
    • Software & computer vision lead: led end-to-end software and computer vision for IoT smart vending across France, Australia, India and Poland, owning the full lifecycle from design and implementation through testing, deployment and production monitoring, scaling to 100+ units per market.
    • Patented invention: invented a Dynamic Product Label device (patent filed), engineering the hardware and firmware that couple YOLO classification with an independent tracking model to project a digital grid mapping products to segmented LCDs.
    • Model fine-tuning & edge acceleration: fine-tuned PyTorch object-detection models (YOLO) for real-time beverage bottle recognition, and optimized and converted the weights with ONNX and OpenVINO to cut inference latency on live edge devices.
    7 more
    • ML validation: replaced an informal test procedure with a statistically rigorous validation suite (99% confidence intervals), and used root-cause analysis on misclassifications to drive retraining and architecture changes, taking classification accuracy from a ~88% baseline to 99% in production.
    • Computer vision: derived a first-principles model for real-time volume calculation with no static reference markers, integrating YOLO and OpenCV stereo vision to estimate volume for moving and partly occluded objects, enabling dynamic pricing on mixed inventory.
    • Hardware-software co-design: architected a depth-based stereo-vision algorithm in OpenCV and C++ to infer door states, replacing error-prone RF sensors; it eliminated field service calls, relaxed OEM assembly tolerances and cut hardware BOM cost by 20%.
    • Edge network routing: designed a resilient LAN topology with an unmanaged switch to ride out 4G dropouts, and a custom NIC update on the Android SoC that assigns subnet-aligned IPs to headless Nayax processors and UI tablets automatically, removing manual ADB configuration and portal provisioning while keeping the WAN port free for hardwired failover.
    • Payment systems: structured a unified payment subsystem for global processors (Nayax, Pax, UPI), with a fault-tolerant pre-authorization and incremental/decremental settlement flow for secure, unattended transactions under intermittent edge connectivity.
    • IoT telemetry: engineered a low-latency NAMA Vending Data Interchange (VDI) telemetry pipeline, with custom firmware that reads the vending machine’s edge device and sends bidirectional hardware states over MQTT, sustaining 99.2% subsystem uptime across 100+ international units.
    • Automated testing: built a black-box and performance testing pipeline (unit, system, integration, load/stress) in Python and ADB, raising test coverage by 30% across firmware, the Android app, cloud and peripheral subsystems.

  3. Undergraduate Researcher

    Student Impact Innovation Lab · NYU Tandon · Medical IoT device (SkiCad)

    Sep 2022 — Jun 2023Brooklyn, NY

    • ESP8266
    • AD5933 · I2C
    • ILI9341 · SPI
    • Firebase
    • Firmware & sensor integration: ESP8266 firmware (Arduino/C++) driving an AD5933 impedance analyzer over I2C, with a calibrated 40-point frequency sweep against a reference resistor, real-time complex impedance (|Z| from the real and imaginary registers) and an interrupt-driven acquisition path.
    2 more
    • Touchscreen UI & cloud pipeline: an ILI9341 TFT touchscreen UI over SPI, streaming measurements over WiFi/HTTP to a Python backend (FastAPI, Flask) with Firebase/Firestore.
    • Visualization: real-time impedance and phase Bode plots (matplotlib, cubic-spline interpolation) for non-invasive skin-impedance characterization.

  4. Undergraduate Researcher

    NYU Tandon Smart Wearable Bio-Tracker · Wearable prosthetic control

    Jun 2022 — Aug 2022Brooklyn, NY

    40×
    Wireless rate · 50 Hz → 2 kHz
    • sMMG
    • UDP
    • Wearable
    • Applied surface mechanomyography (sMMG) with differential accelerometer readings to capture muscle signals for real-time hand-movement prediction.
    1 more
    • Raised the wireless transmission rate 40× (50 Hz → 2 kHz) by migrating to UDP, enabling responsive prosthetic control.
U10 · EDU-ROMEducation

Burned into memory

  1. New York University

    Tandon School of Engineering

    2021 — 20230x00 · Bank 0

    GPA 3.8 / 4.0

    B.S. in Computer Engineering · Minor in Science & Technology Studies

    • Cum Laude
    • 2× Dean’s List

    Coursework13 cells

    • 0x00Computer Architecture
    • 0x01Machine Learning
    • 0x02Operating Systems
    • 0x03Databases
    • 0x04Digital Logic & Machine Design
    • 0x05Real-Time Embedded Systems (Grad)
    • 0x06Embedded Systems
    • 0x07VLSI Circuits
    • 0x08Fundamentals of Electronics
    • 0x09Advanced Linear Algebra
    • 0x0AComplex Variables
    • 0x0BDiscrete Math
    • 0x0CData Analysis

  2. Brooklyn College

    Department of Engineering

    2019 — 20210x01 · Bank 1

    GPA 3.7 / 4.0

    Computer Engineering

    • Engineering Honors
    • Dean’s List

    Coursework12 cells

    • 0x00Intro Programming (Java)
    • 0x01Object-Oriented Programming
    • 0x02Data Structures
    • 0x03Algorithms
    • 0x04Paradigms in C++
    • 0x05Engineering Physics I–II
    • 0x06Electric Circuits
    • 0x07Electrical Measurements Lab
    • 0x08Engineering Mechanics
    • 0x09Calculus I–III
    • 0x0ALinear Algebra
    • 0x0BDifferential Equations
Projects
U11 · LIVE-LION01 / 10

The client platform I built for Live Lion Security, a remote video monitoring company. Live at clients.livelionsecurity.com, it replaced scattered per-site WhatsApp groups with one web app for alarms, escalations and camera setup.

Live
client platform
Pilot
rolling out to client sites
Built to scale
multi-site architecture
clients.livelionsecurity.com Case study
U1 · BP-MONITOR02 / 10

A bare-metal blood-pressure monitor on an ATmega32U4 (Adafruit Circuit Playground). It reads a Honeywell MPRLS pressure sensor over I2C and extracts systolic, diastolic and heart rate from the cuff’s pressure waveform.

5-stage
bare-metal state machine
24-bit
raw sensor telemetry
20-sample
peak-detection buffer
Case study

Interactive demo: Pump the cuff and read the pressure curve

U2 · SENTRY03 / 10

A multi-threaded, event-driven gesture lock on an STM32 with Mbed OS. An interrupt-driven 1 MHz SPI pipeline reads the gyroscope, and Dynamic Time Warping matches the gesture at any speed.

1 MHz
interrupt-driven SPI
DTW
variable-speed matching
Case study

Interactive demo: Draw a gesture to set the key, then unlock it

U3 · CV-SIZE04 / 10

At PepsiCo, I derived a first-principles model for real-time volume calculation with no static reference markers. YOLO and OpenCV stereo vision estimate the volume of moving, partly hidden products, which enables dynamic pricing on mixed inventory.

0
reference markers needed
Real-time
moving, occluded objects
Case study

Interactive demo: Drag a product and watch stereo estimate its size

U4 · DOOR-DET05 / 10

A depth-based stereo-vision algorithm in OpenCV and C++ that infers a vending machine’s door state, replacing error-prone RF sensors. It eliminated field service calls and cut hardware BOM cost by 20%.

20%
hardware BOM cost cut
RF → depth
sensor replaced
Case study

Interactive demo: Open the door and watch the depth map decide

U5 · FOREX-ENG06 / 10

A 24/7 automated signal-detection system in Python on the Oanda v3 REST API. It watches 68+ currency pairs concurrently and batches every update into one Telegram broadcast, running on AWS EC2 under PM2.

68+
pairs monitored at once
24/7
on AWS EC2
Case study

Interactive demo: Feed simulated ticks to the producer threads

U6 · CANCER-ML07 / 10

Survivability prediction on clinical staging data. I compared logistic regression, SVM and neural-network models, and the best one raised test accuracy from a 78.5% baseline to 91.3%.

91.3%
test accuracy (from 78.5%)
3
model families
Case study

Interactive demo: Move the clinical features, run the model live

U7 · AIR-TICKET08 / 10

A dual-interface flight reservation system in Python Flask on a normalized MySQL database, with separate customer and airline-staff portals and reporting built on aggregate SQL.

2
portals: customer and staff
Normalized
MySQL schema
Case study

Interactive demo: Book a fake flight through both portals

U8 · E20-CPU09 / 10

A two-pass E20 assembler, an object-oriented instruction-set simulator for the full ISA, and a configurable two-level cache simulator, all in C++ and validated against 40+ test scenarios.

40+
test scenarios
L1/L2
cache with LRU
Case study

Interactive demo: Step a tiny E20 CPU through real assembly

U9 · SMART-SHADES10 / 10

Custom ESP32 firmware that automates residential shades: H-bridge PWM control (TB6612FNG) of high-torque worm-gear motors with sub-100 ms response, plus Tuya IoT cloud control, schedules and manual overrides from a phone.

<100 ms
response latency
Tuya
cloud control
Case study
J2 · PIN HEADERSkills

Every pin I can drive

  • LANG
    LanguagesC/C++ · Python · Java
  • OS
    Operating systemsBare-Metal · Embedded Linux · FreeRTOS
  • PROTO
    ProtocolsEthernet · TCP · UDP · I2C · SPI · UART · RS-232 · RS-485 · MQTT · Tuya · REST · WebSocket · HTTP/HTTPS
  • MCU
    MCUs & SoCsATmega · Arduino · ESP32 · STM32 · Raspberry Pi · Qualcomm Kryo 670 SoC (Android)
  • ML
    Machine learningPyTorch · TensorFlow · Keras · Scikit-learn · YOLO · OpenCV · ONNX · OpenVINO · TensorRT · NumPy · Pandas · Matplotlib
  • CLOUD
    Backend & cloudFastAPI · Flask · Django · AWS (EC2, S3) · PostgreSQL · MySQL · MongoDB · SQLite · Firebase · Supabase · Cloudflare
  • DEBUG
    Tools & debuggingGit · GitHub Actions · pytest · unittest · Jupyter/Colab · PM2 · ADB automation · GDB · Log analysis · Logic analyzers · Oscilloscope
AI-generated voice and avatar
Runs entirely in your browserCase study