Research on spiking neural networks and low-power on-device AI.
Akran IQ is an independent research effort exploring event-driven neural networks, neuromorphic compute, and ultra-low-power inference for the next generation of edge devices. We publish notes, primers, and experiments — not products.
What we’re researching
Four overlapping threads. Most posts touch more than one — the interesting questions sit at the boundaries.
Spiking Neural Networks
Event-driven models, surrogate-gradient training, and where SNNs actually beat dense ANNs on energy and latency budgets.
Low-Power On-Device AI
Quantization, pruning, and architecture choices for sub-milliwatt inference on microcontrollers and edge accelerators.
Neuromorphic Compute
Notes on Loihi, SpiNNaker, Akida, and the trade-offs between event-driven silicon and conventional NPUs.
On-Device Learning
Continual learning, federated updates, and what it takes to train — not just infer — on resource-constrained hardware.
Latest writing
Recent notes from the blog. Older posts cover practical IoT and edge-device topics that informed the research direction.
LoRaWAN vs NB-IoT in 2026: A Practical Decision Guide for IoT Deployments
LoRaWAN and NB-IoT are the two dominant LPWAN technologies in 2026. This guide compares them on cost, coverage, battery life, latency, private network control, and total cost of ownership, with clear recommendations for smart metering, agriculture, logistics, and industrial IoT.
Read more →
Fiber-Optic Drones Explained: Why Militaries Are Ditching RF Control Links
Fiber-optic drones use ultra-thin optical fibers instead of radio waves to transmit control signals and HD video. Immune to jamming, low-latency, and high-bandwidth, they are reshaping battlefield surveillance and industrial tethered operations in 2026. Here is how the technology works and where it fits.
Read more →
Bluetooth Mesh vs Thread vs Matter in 2026: When to Choose What
Bluetooth Mesh, Thread, and Matter are often lumped together but they solve different problems. This guide breaks down what each actually is, where they overlap, and how to pick the right stack for smart homes, commercial buildings, and industrial IoT deployments in 2026.
Read more →
Convert PNG Gas Meters to Smart Meters at Scale: LPWAN vs WiFi Mesh vs Bluetooth Mesh for Gas Utilities
City gas distribution companies and housing societies can convert existing PNG diaphragm meters into smart gas meters using wireless networks. This guide compares LPWAN (LoRaWAN), WiFi mesh, and Bluetooth mesh on safety, cost, latency, and scalability for society-level gas meter reading in India.
Read more →About this site
Akran IQ started as a working notebook for problems we kept hitting at the edge: models that were fast enough on a server but melted a microcontroller, accelerators that looked great in a datasheet and terrible under real workloads, and a quiet sense that the standard dense-ANN recipe is not where the next order-of-magnitude improvement in efficiency comes from.
The current focus is spiking neural networksand event-driven inference — what they’re actually good for, what training them honestly costs, and how close neuromorphic hardware is to being the right tool for production edge AI rather than a research curiosity.
Posts are written when something is worth writing down, not on a schedule. If a topic looks interesting and is missing, send a note.
Frequently asked questions
What is Akran IQ?
Akran IQ is an independent research effort focused on spiking neural networks, neuromorphic compute, and ultra-low-power on-device AI. The site publishes research notes, primers, and experiments — it is not a product or service.
What are spiking neural networks (SNNs)?
Spiking neural networks are event-driven neural models inspired by biological neurons. Instead of dense matrix multiplications on every step, neurons fire discrete spikes only when their internal state crosses a threshold. This sparsity is what makes SNNs interesting for low-power edge AI on neuromorphic hardware like Intel Loihi, BrainChip Akida, and SpiNNaker.
How is this different from standard TinyML?
TinyML usually means running quantized dense neural networks on microcontrollers. SNN and neuromorphic research asks a different question: what if the model itself is event-driven, so most neurons do nothing on most timesteps? Our writing covers both — TinyML is the practical baseline; SNNs are the research bet on the next order-of-magnitude efficiency gain.
Do you offer consulting or build products?
No. The site is a research notebook. For collaborations, citations, or topic suggestions, use the contact page.
Where do older IoT posts fit in?
Earlier posts on LoRaWAN, NB-IoT, Bluetooth Mesh, CAN bus, and IoT platforms remain published because the constraints they describe — tiny power budgets, intermittent connectivity, real-time deadlines — are exactly the constraints that make event-driven neural compute interesting in the first place.