</> Manas Punjabi
Resume

Embedded Firmware · ECE

Manas Punjabi

Embedded Firmware Engineer

I build systems where software meets hardware — microcontrollers, RTOS, wireless protocols and connected devices that have to survive the physical world.

This page is a journey. Scroll through it.

01

Where it started

Mumbai, India · Dwarkadas J. Sanghvi College of Engineering

B.S. Electronics Engineering · May 2021

University of Mumbai

Four years of electronics fundamentals, weighted hard toward embedded systems — microcontrollers, RTOS, digital design and communication. The groundwork for everything that came after.

  • Microcontrollers & Applications
  • Embedded Systems & RTOS
  • Digital System Design
  • Digital Signal Processing
  • Internet of Things
  • Computer Organization & Architecture
  • Digital Communication
  • Instrumentation System Design

DJ Sanghvi Formula Student · Electronics Dept.

Firmware at 8,000 RPM

My interest in embedded systems started well before my first job — on a Formula Student race car, where timing, packaging and reliability were not abstractions. If the firmware was late, the shift was late.

  • 0 faster gear shifts — STM32-controlled pneumatic actuators
  • 0 reliability gain — from a redesigned electrical harness
  • 1st place, Cost & Manufacturing Event — Red Bull Ring, Austria
  • STM32
  • Embedded C
  • Actuator Control
  • Harness Design

Academic Project · DJ Sanghvi

Crash Detection & Emergency Alert

An Arduino system that watched accelerometer data for the abrupt motion change that marks a crash, then dispatched an emergency notification carrying GPS coordinates over GSM — to emergency services and to family.

  • Auto no user action required after impact
  • GPS coordinates embedded in the alert
  • Arduino
  • Accelerometer
  • GPS
  • GSM
  • Embedded C
MCU

Fundamentals

SHIFT

STM32 · pneumatic shift control

IMPACT GPS FIX ! CRASH DETECTED COORDS SENT

Detect → locate → notify

02

Bar Code India

Embedded Firmware Engineer · RFID, BLE, IoT & asset tracking · Pune, India · Mar 2022 – Sep 2023

RFID · Connectivity

Automatic Toll Collection Reader

The reader’s STM32 turned tag reads into JSON. My half started there — an ESP32 that ingested that stream, buffered it, and got it onto whatever network the site had: sockets, MQTT and AWS over WiFi, GSM or Ethernet.

  • 0 tag reads ingested and forwarded
  • 0 read success in test scenarios
  • 3 interchangeable transports, switchable by config
  • ESP32
  • MQTT / AWS
  • WiFi / GSM / Ethernet
  • RTOS Tasks
  • State Machine

BLE · Sole owner

Low-Power Beacon & Gateway

My first project carried from blank page to finished product — and mine end to end, both the beacon and the gateway. Nordic nRF52 under Zephyr RTOS, running off a coin cell, built for zonal tracking, mobile app integration and wireless asset telemetry.

  • Solo beacon and gateway firmware, start to finish
  • Coin cell — the entire power budget
  • Zonal tracking, app integration, asset telemetry
  • nRF52
  • Zephyr RTOS
  • BLE
  • ESP32
  • Low Power

Localization · Sole owner

BLE Asset Tracking Gateway

Static beacons placed around an office, with the gateway resolving asset position from RSSI to keep a live zonal picture of where things were. Mine end to end, building directly on the beacon and gateway firmware before it.

  • Solo designed and built start to finish
  • 0 static asset positioning accuracy
  • Zonal live positions held across the office
  • BLE
  • RSSI
  • Trilateration
  • Static Beacons
  • Zonal Positioning

Wearables · My first project here

Wearable Pedometer

The BLE stack and the app already existed, so I used this as my way into sensor integration — turning raw accelerometer output into a step count that held up for everyone, not just for me, and streaming it live for labor fatigue monitoring.

  • 0 step-detection accuracy
  • 5,000 steps walked by each volunteer to find a shared threshold
  • R&D built and validated, never taken to market
  • Accelerometer
  • BLE
  • Thresholding
  • Signal Analysis
  • Wearable

Industrial · Condition Monitoring

Machine Vibration Monitoring

The same accelerometer, pointed at machines instead of people. The device watched vibration in an industrial environment, and when a machine crossed threshold enough times it was rated good, moderate or bad on a dashboard — a read on machine health without an engineer standing next to it.

  • 3 health states from repeated threshold crossings
  • Same sensor work carried over from the pedometer
  • Tested developed and validated in house
  • Accelerometer
  • Vibration
  • Condition Monitoring
  • Thresholding
  • Dashboard
READER TAGS / SEC 200

~200 tags/s · 99% accuracy

nRF52 ZEPHYR

nRF52 · Zephyr RTOS · BLE

~5 m

RSSI trilateration · ~5 m accuracy

g t threshold STEP ACCURACY 96%

Peak detection · 96% accuracy

GOOD MODERATE BAD MACHINE HEALTH VIBRATION

Threshold crossings → health rating

03

Frinso Technologies

Sole embedded developer in a 5-person team · Mumbai, India · Mar 2024 – Jul 2024

OTA

Over-the-Air Firmware Updates

Deployed devices could take new firmware remotely — no physical recall, and firmware could be customized after the device was already in the field.

  • 0 device recalls needed for firmware changes
  • Field post-deployment customization
  • ESP32
  • OTA
  • C / C++
  • WiFi

Sensor Integration

ESP32 Modbus Sensor Hub

One hub, many sensors, speaking Modbus and UART — chlorine detection, humidity and whatever else the deployment called for — plus a WiFi access-point webpage so a user could pick and configure sensors with minimal setup.

  • Cl₂ chlorine, humidity and other Modbus sensors
  • AP browser-based sensor configuration on site
  • ESP32
  • Modbus
  • UART
  • RS-485
  • Web Config

Power

Deep Sleep Power Optimization

The hub ran on batteries, so current draw was the design constraint. I put the device to sleep between acquisitions — then measured the difference the honest way: run it on a battery until it died, with deep sleep and without, and compare how long it lasted.

  • 0 longer battery life
  • Bench measured on battery, not estimated
  • Deep Sleep
  • Low Power
  • Battery
  • ESP32

Reliability

SD-Card Data Fallback

Networks are not always there. When the device had no connection, readings went to local SD storage instead of disappearing, so an outage cost you a delay rather than a hole in the data.

  • 0 readings lost during network interruptions
  • SD local storage whenever the network is down
  • SD / SPI
  • Fault Tolerance
  • Buffering
  • ESP32

Alerting · Safety

River Level Early Warning

An ESP sensor hub and a SIMCOM GSM module watching river levels as a safety system. When the water crossed the danger line it sent an SMS straight to the responsible authorities at the company — no app to open, no internet required.

  • Alert fires the moment the level crosses the danger line
  • SMS reaches a phone where there is no internet
  • ESP32
  • SIMCOM GSM
  • SMS
  • River Level
  • Safety System
FIRMWARE ESP32 UPDATING

Remote firmware delivery

ESP32 MASTER CL2 RH S3 MODBUS / RS-485

One master, many sensors

mA always-on z z z +50%

Measured on battery, sleep vs. always-on

HUB SERVER SD BUFFERED

Link fails → data survives

DANGER LEVEL ESP + GSM RIVER LEVEL ! RIVER LEVEL CRITICAL SMS SENT TO AUTHORITIES

Crosses the line → SMS goes out

04

8,000 miles

Mumbai, India  →  Chico, California

MUMBAI, IN CHICO, CA

Four years of product firmware behind me.

A graduate program ahead.

One very long flight in between.

05

California State University, Chico

M.S. Electrical & Computer Engineering · 3.7 GPA · Expected December 2026

Graduate Study

From registers to representations

Graduate work widened the stack: the same systems I used to build at the firmware level, now viewed through signal processing, learned models and secure communication. Low-level embedded understanding on one end, data and security on the other.

  • Advanced Embedded Systems
  • Embedded Digital Design
  • Digital Signal Processing
  • Machine Learning
  • Deep Learning
  • Secure Computing

2026 · Firmware + IoT

Industrial Hazard Sensor & Dashboard

Six sensor and peripheral driver modules in Embedded C on a Tiva TM4C123, feeding a menu-driven OLED UI and an MQTT dashboard.

  • 6 driver modules, each with its own API
  • 6 gases — CO, NO₂, NH₃, CH₄, H₂, ethanol
  • Fixed point maths — no FPU on this part
  • Tiva TM4C123
  • TivaWare
  • Embedded C
  • I2C
  • ADC
  • MQTT
  • Raspberry Pi

2026 · Security · IoT

Secure Telemetry over ECIES and TLS

A C client and server carrying a JSON temperature reading from a sensor node to a collector under two independent layers of encryption — TLS for the transport, and an ECIES envelope around the payload itself, sealed with a fresh key for every message.

  • 2 independent layers — defence in depth
  • Per message ephemeral key — forward secrecy
  • P-256 ECDH → HKDF-SHA-256 → AES-256-GCM
  • C
  • mbedTLS
  • TLS
  • ECDH
  • HKDF-SHA-256
  • AES-256-GCM
  • Base64

2025 · VLSI Design & Verification

Matrix Multiplier in SystemVerilog

A parameterized matrix multiplier where every element-wise product gets its own pipelined shift-and-add multiplier, all running in parallel, sequenced by a five-state FSM — the hardware version of arithmetic firmware usually does in software.

  • 6.7× throughput gain over the non-pipelined design
  • 38 cycles latency, down from ~260
  • 416 MHz, closed with 11 ps of setup slack
  • SystemVerilog
  • Pipelining
  • FSM
  • ModelSim
  • Xilinx FPGA
signal FFT SAME SYSTEM, TWO ALTITUDES FIRMWARE LEARNING

DSP · ML · Secure Computing

TM4C123 GPIO ADC I2C SYSTICK DHT11 MPU6050 GAS SOUND MQTT DASHBOARD ! HAZARD ALERT

Bare-metal drivers → live dashboard

CLIENT SENSOR SERVER COLLECTOR EPHEMERAL ECDH PUBLIC KEY CERTIFICATE-VERIFIED TLS HKDF-SHA256 SHARED SECRET → SESSION KEY AES-256-GCM

Ephemeral keys · authenticated encryption

A [4x4] × B [4x4] = C [4x4] PIPELINED + PARALLEL · 6.7× THROUGHPUT TIMING CLOSED AT 416 MHz

Parameterized RTL · pipelined & parallel

06

Master’s Thesis

AI-Driven Learning from Complex Multitemporal Satellite Image Data · Advisor: Dr. Hassan S. Salehi · CSU Chico

Yolo County, California

Telling crops apart from orbit

Crop information still comes largely from grower surveys and county reports — costly, labour-intensive and slow to update. The thesis builds the automated alternative: classify Almond, Alfalfa and Uncultivated AG parcels straight from Sentinel-2 imagery, in collaboration with the College of Agriculture.

  • 100k parcel images across 2017–2024
  • 10 m resolution, 5-day satellite revisit
  • 3 classes, balanced to 16,146 images each
  • Sentinel-2
  • Google Earth Engine
  • Python
  • MATLAB

Phase I

Hand-built features, Random Forest

Vegetation indices and statistical features extracted per parcel, then three feature-selection algorithms — mRMR, ReliefF and a Genetic Algorithm — ranked which of them actually carried information before a Random Forest used them.

  • 83.6% best testing accuracy (ReliefF)
  • 3 selection algorithms compared head to head
  • Random Forest
  • mRMR
  • ReliefF
  • Genetic Algorithm
  • Vegetation Indices

Phase II

Four CNNs, transfer learning

VGG-16, VGG-19, ResNet-50 and GoogLeNet fine-tuned on the parcel imagery, with augmentation — ten rotations plus flips — growing the training set to roughly 140,000 images per class.

  • 98.8% VGG-19, random split
  • 4 architectures compared across three splits
  • VGG-16
  • VGG-19
  • ResNet-50
  • GoogLeNet
  • Transfer Learning

The result that matters

The number falls when the test gets honest

A random split lets images from the same farm and the same year sit in both training and test sets, which flatters the model. So I also split by farm — can it handle fields it has never seen? — and by year — can it handle an unseen season? Accuracy drops, and that drop is the finding.

  • 98.8% random split — flattering
  • 85.4% farm split — spatial generalization
  • 83.7% year split — temporal generalization
  • Evaluation Design
  • Spatial Generalization
  • Temporal Generalization

Phase III

Putting it in front of a user

The Random Forest and CNN models integrated into a web interface for crop classification, so the workflow ends somewhere a grower or an analyst can actually use — not in a notebook.

  • End to end: acquisition → classification → map
  • Both model families served from one interface
  • Web Interface
  • Python
  • Model Serving

Publications

Published from this research

  • AI-powered learning from complex satellite image data: integrating random forests and deep learning for crop classification and assessment

    Manas Punjabi, Hassan S. Salehi, Hossein Zakeri

    Proc. SPIE 14030, Machine Learning from Challenging Data 2026, 140300S · 12 June 2026

    DOI: 10.1117/12.3095166 →

Supported by the Professor Gee Memorial Award, with equipment partially funded through NSF MRI-1920345.

SENTINEL-2 · 10 m · 5-DAY REVISIT PARCEL-LEVEL NRG COMPOSITES PARCEL IMAGES 100,495 2017 – 2024

Yolo County · 8 seasons

RANKED FEATURES mRMR · ReliefF · GA RANDOM FOREST BEST TESTING ACCURACY 83.6%

Select the features, then classify

TRANSFER LEARNING RANDOM SPLIT ACCURACY VGG-19 98.8 VGG-16 98.7 ResNet-50 98.1 GoogLeNet 96.6 AUGMENTATION: 10 ROTATIONS + 2 FLIPS ≈ 140,000 TRAINING IMAGES PER CLASS

Four architectures, one dataset

98.8 RANDOM 85.4 FARM 83.7 YEAR MORE REALISTIC EVALUATION →

Unseen fields · unseen seasons

ALMOND ALFALFA UNCULT. MODEL CNN / RF CROP CLASSIFICATION WEB INTERFACE

Models, served

SPIE 14030

Peer-reviewed

07

The toolbox

What I reach for

Languages

  • C
  • C++
  • Embedded C
  • Python
  • SystemVerilog
  • MATLAB
  • JavaScript

Firmware & RTOS

  • Bare Metal Firmware
  • Peripheral Drivers
  • Board Bring-Up
  • FreeRTOS
  • Zephyr RTOS
  • Tasks / Queues / Semaphores
  • Interrupts & ISRs
  • Timers
  • Low Power Modes
  • OTA & DFU
  • Firmware Validation

MCUs & Platforms

  • STM32
  • ESP32
  • nRF52
  • Tiva TM4C123
  • Raspberry Pi
  • Arduino
  • Xilinx FPGA

Interfaces

  • I2C
  • SPI
  • UART
  • RS-232
  • RS-485
  • Modbus
  • USB
  • BLE
  • NFC
  • WiFi
  • Ethernet
  • MQTT
  • TCP/IP
  • GPIO
  • ADC
  • PWM

Debug & Tools

  • JTAG & SWD
  • Logic Analyzer
  • Oscilloscope
  • Git
  • STM32CubeIDE
  • nRF Connect SDK
  • TivaWare
  • Keil
  • Eclipse
  • VS Code
  • ModelSim
  • Eagle
  • Proteus

Security & Cloud

  • mbedTLS
  • TLS
  • ECIES
  • ECDH
  • AES-256-GCM
  • HKDF-SHA256
  • FastAPI
  • WebSocket
  • HiveMQ
08

What’s next

Graduating December 2026

I’m looking for full-time work in embedded firmware, real-time systems, device drivers, connected devices, robotics and automotive — the places where firmware and hardware have to agree with each other. If that’s your team, I’d like to hear from you.