Modern technology is built from layers that continuously interact with one another. At the physical level, electronic devices transform electrical signals into digital information. Digital logic turns that information into computation, while processors, memory and communication interfaces provide the machinery needed to execute instructions and move data. As these components become part of embedded systems, they begin to interact with the physical world through sensors, controllers, actuators and real-time software.
But computation does not exist in isolation. Operating systems coordinate hardware and software, firmware gives specialized machines their behaviour, and communication protocols allow independent systems to exchange information. At the same time, machine learning is moving beyond the cloud into edge and on-device systems, where models must operate within real constraints such as memory, processing power, latency and energy consumption.
PrajnaEdge explores these connections as one continuous technology landscape — and takes them beyond explanation. From computing foundations and embedded systems to intelligent machines and edge AI, ideas can be understood, experimented with, and eventually turned into technology that can be experienced in the real world.
Experiment with intelligence beyond the cloud.
Explore how neural networks learn representations from data before compression and deployment. Future experiments will allow adjusting learning parameters—such as learning rate, epoch counts, and batch sizes—to observe loss trajectories and decision boundaries.
Can this image classifier maintain its intelligence while becoming small enough for the edge?
Run a fully quantized INT8 image-classification model directly on embedded hardware.
Can this image classifier maintain its intelligence while becoming small enough for the edge?
Run a fully quantized INT8 image-classification model directly on embedded hardware.
This experiment investigates preparing a compact INT8 TensorFlow Lite model for direct inference on resource-constrained embedded hardware. The model uses the same CNN architecture as the Edge AI exploration, converted to a full INT8 representation to classify three classes: Apple, Banana, and Orange.
Explore the ideas, systems and connections that shape technology — choose any node to begin your journey.
Practical guides to embedded systems and engineering.
Deploying neural networks and intelligent decision loops on raw silicon targets.
How linker scripts quietly decide where computation lives.
To a programmer working in a high-level application space, memory is often visualized as a flat, infinite landscape of bytes—a continuous array indexed from zero to the limits of physical RAM. You allocate memory, declare variables, and write functions under the assumption that all space is structurally equivalent.
But beneath the abstraction layers of low-level firmware, this uniformity disappears. Memory is not a single contiguous pasture; it is a highly structured, compartmentalized terrain. A microcontroller's memory space is fragmented by physical and electrical realities. Code, initialized global variables, uninitialized flags, read-only constants, stack frames, and dynamic heaps must reside in completely different areas of the silicon to function.

Firmware is not stored as one object. It is distributed across memory according to behavioral requirements.
Why do we slice firmware into segments? It is a direct response to different behavioral requirements. Different classes of information behave differently during execution, requiring distinct storage mediums.
Consider instructions: executable code must be persistent and read-only, which is why the compiler assigns it to the '.text' section to reside in Flash memory. Constants and string literals share similar read-only behaviors and are grouped into the '.rodata' section, protected from runtime modification.
Global variables, however, present a dual behavior. They must be modifiable at runtime (residing in RAM), but they must start with a pre-configured initial value. These are grouped into '.data'. Uninitialized global variables—those that simply need to be cleared to zero on startup—are routed to '.bss' (Block Started by Symbol) to save space in the flash binary.
Finally, the Stack handles execution flow by dynamically pushing and popping function parameters and local variables, while the Heap manages runtime dynamic memory allocation workspace. The organization of these sections is the layout of digital life.
During compilation, the compiler translates individual source files (.c or .cpp) into independent object files (.o). At this stage, the compiler acts in isolation: it has no knowledge of how much total Flash or RAM exists on the chip, nor does it know where other object files will eventually be placed.
This is where the Linker enters as a cartographer. The linker compiles these fragmented objects and maps their sections into a unified, physical landscape. It decides the precise layout of the firmware—stitching together the various '.text', '.data', and '.bss' fragments and matching them to the exact address boundaries of the target processor.

The compiler generates isolated segments; the linker script guides the linker to map them into physical memory blocks.
To direct the linker, we use a linker script—a blueprint that defines the physical boundaries of the MCU's address register space. It tells the linker exactly where Flash and RAM begin and how long they are.
In this snippet, we define the region boundaries. 'FLASH' is marked as read-only and executable (rx) starting at origin address 0x08000000 with a length of 512 Kilobytes. 'RAM' is marked as readable, writable, and executable (rwx) starting at origin 0x20000000 with 128 Kilobytes.
By establishing these constraints, the compiler can output absolute addresses, knowing that the code and variable segments will fall safely within the silicon boundaries.

Microcontroller address registers are mapped to specific Flash (executable) and RAM (read-write) regions.
When power is applied, the microcontroller is in a primitive state. The RAM is filled with random electrical noise. However, global variables in the '.data' section must have their initial non-zero values ready for the application to read.
Since RAM is volatile and loses all state when powered off, these initial values must be stored persistently inside Flash memory. This creates a logical conflict: the variables must be modifiable in RAM during execution, but their initial values are trapped in Flash.
To resolve this, the system executes startup assembly or C initialization routines before your main() function is reached. This startup code performs the first migration: it reads the block of initial values from Flash and copies them into their RAM destinations. It then clears the '.bss' section in RAM to zero.

Relocation sequence at startup, moving initialized variables from Flash to RAM and zeroing out the BSS.
Once the linker has completed its map, it outputs a map file—a census of the compiled firmware. The map file outlines exactly where every function, variable, and library symbol has been placed, along with the size of each segment.
Experienced embedded engineers inspect map files regularly. When a firmware image suddenly overflows memory, the map file reveals which module or library is consuming the flash. It lets you identify compiler optimizations, alignment padding issues, and variables that were accidentally declared globally instead of locally.
Memory mapping eventually moves from a compiler detail to a system design challenge. In advanced architectures, you must place DMA (Direct Memory Access) buffers into specific memory regions that the DMA controller can actually access. If the buffer is mapped to a core-coupled RAM region that is inaccessible to the peripheral bus, the system will silently fail or crash.
Similarly, separating a bootloader from an application requires carving Flash memory into separate, non-overlapping regions using custom linker scripts. The developer is no longer writing software alone—they are shaping the physical layout of logic.
Memory architecture defines the boundaries of what is electrically possible, but the linker script defines the geography where firmware actually lives.
As you write embedded code, you are not just writing logic; you are organizing silicon.
PrajnaEdge is a technology company exploring the space between understanding technology, experimenting with ideas, and turning them into things that can be experienced.
PrajnaEdge began with Embedded Systems — exploring the foundations that connect hardware, software and intelligent computation.
The first technology universe is built around that foundation. The journey will expand as new ideas, experiments and products emerge.
PrajnaEdge is a technology company created by Devaharsha Meesarapu.
I am the engineer behind the design, development, and content of PrajnaEdge. I build low-level systems where code directly controls hardware, bridging the gap between register-level silicon behavior and intelligent edge decision loops.
I am an Embedded Firmware Engineer focused on developing software for resource-constrained systems. My experience spans bare-metal firmware, device drivers, microcontroller peripherals, and communication protocols, working across the boundary between hardware and software.
My work has involved microcontroller-based systems, real-time behaviour, hardware interfaces, and communication technologies such as CAN, CAN FD, UART, SPI, and I²C. I am particularly interested in understanding systems from the lowest level upward—from registers and peripherals to intelligent edge systems.
Engineering is not just about writing code; it is about managing constraints, timings, and physical hardware characteristics. True mastery of complex systems comes from understanding the interactions across different layers of the stack.
This conviction is why I built PrajnaEdge—to bridge the gap between conceptual theory and direct, register-level physical reality.
Software that runs directly on hardware without an operating system.
"Every embedded application begins long before main()."
An Operating System manages hardware and software resources so complex applications can work efficiently.
"When one loop is no longer enough to carry the burden."
PrajnaEdge is an independent education platform built to make knowledge freely accessible.
If you find PrajnaEdge useful, you can support its continued development.
Your support helps fund the time, tools, infrastructure, and experimentation that go into building and maintaining PrajnaEdge.
Welcome to PrajnaEdge (prajnaedge.dev), an independent engineering and technology platform created and maintained by Devaharsha Meesarapu. By using this website, you agree to these terms.
Educational & Research Focus: PrajnaEdge publishes interactive technical explorations, architectural models, and simulation walk-throughs covering embedded systems, computer architecture, operating systems, and edge artificial intelligence. All materials, interactive tools, and code demonstrations are provided solely for educational and conceptual understanding.
Hardware & Firmware Disclaimer: Embedded programming interacts directly with hardware registers, physical voltages, and precise timing. While all writeups and code demonstrations are prepared with care, they are provided "as is" without warranty of any kind. You are responsible for reviewing component datasheets, circuit schematics, and electrical ratings before deploying code to physical microcontrollers or custom hardware.
Intellectual Property: All original articles, custom SVG architectures, interactive simulators, curriculum sequences, and source code are the intellectual property of Devaharsha Meesarapu (© 2026 PrajnaEdge. All rights reserved). Non-commercial educational study and citation are welcome with proper attribution. Direct republication, unauthorized mirroring, or mass scraping of content is not permitted.
PrajnaEdge is committed to user privacy and minimal data collection. We do not sell, rent, or monetize your personal information.
localStorage purely to remember your interface preferences on your device (such as sound preferences for animations and tutorial display states). No personal identity data is stored in localStorage.
To help support platform operation and hosting, PrajnaEdge displays advertisements served by third-party advertising partners, including Google AdSense.
We use Google Analytics (measurement tag: G-6Y8ZVQB1V0) to evaluate anonymous, aggregate usage trends across our technical writeups.
Depending on your jurisdiction (including rights under GDPR, CCPA/CPRA, and applicable privacy laws), you have the right to request access to, correction of, or deletion of any personal communications you have submitted. We do not sell personal data.
For any questions regarding these terms, privacy practices, or data inquiries, contact the creator directly: