Polarj Sapkota
Electrical Engineer · Theorist
Electrical Engineer · Theorist
If life is a latent space, movement is the only reward.
Hi I'm Polarj, an electrical and electronics engineer with specializations in power and control systems from Kathmandu University. I've built wireless sensor networks for wildlife conservation, shimmed superconducting magnets for MRI systems at General Electric, designed flight controllers for eVTOL aircrafts, and written compilers that turn natural language into verified code.
The thread connecting these is a persistent interest in systems under constraint, specifically—energy, size & SNR—and in building tools that remove friction between thinking and building.
My current life is desingned around maximizing theoretical and applied experimentation. My research direction centers on nanosystems and nanodevices: multi-modal energy harvesting, far-field wireless power transfer, and monolithic integration at the nanoscale. The goal is creating biomimetic devices like the microbots you saw in Big Hero 6 with an even smaller footprint.
I've also recently taken up some interest in compilers after taking a rigorous tour through CS. I'm currently working towards building a natural language compiler that increases LLM-inference efficiency by a factor of ~100,000 compared to current state-of-the-art.
Major intellectual contributions and active research directions.
A framework for making LLM-generated code deterministic and verifiable. Each algorithmic step is specified as a domain–range contract: the LLM generates code, the system asserts correctness, and regenerates on failure. The result is a compiler that accepts natural-language algorithm descriptions and produces verified, working programs.
The thesis that all computational behavior reduces to ~650 structural primitives—analogous to the periodic table's 118 elements yielding millions of compounds. Working prototype: 35 indexed fragments with domain/range contracts, deterministic SHA-256 caching, and 5 orders of magnitude energy savings (2 J/token compile vs. 36 μJ/token serve).
A book-length research essay analyzing 260 years of labor transitions through the lens of AI-driven automation. Central finding: when a capability becomes cheap, labor migrates one abstraction layer up. Examines SWE-bench benchmarks, Cursor's growth, junior developer employment data, and the 4.8M unfilled cybersecurity roles.
Designed and prototyped sensor nodes using 4-element microphone arrays with 2° angular discrimination, spectrogram-based noise removal, and novel PSD estimation via abstract algebra. Detects poaching activity (gunshots, chainsaws, engines) in real-time across protected habitats. Validated by four major conservation organizations.
R&D system calculating real-time ampacity based on air/conductor temperature, wind speed, and solar heating. Includes predictive models for ampacity forecasting to enable dynamic line rating. Published at IEEE PowerCon 2024.
ANN & LSTM-based predictive models for detecting abnormal cardiac conditions including atrial fibrillation. Trained on ECG time-series data with real-time classification output.
Hardware, power systems, control, instrumentation, and design projects. Click thumbnails to explore.
Click any image to enlarge.
Physical builds: microphone arrays, thermal rating apparatus, circuit boards.
RGB Art Gallery lighting renders — professional illumination design with colored zones.
Edge detection operators, histogram equalization, image negatives, spectral analysis, and complex field analysis.
Collection of analog, digital, and mixed-signal circuit schematics.
From compilers to quantum circuits to AI-powered entertainment.
Live compiler: write algorithms in plain English, get verified working code. Iterative domain–range assertion ensures deterministic outputs from a probabilistic LLM.
35 indexed computational fragments with contracts, 5 composition pipelines, deterministic cache. Proof-of-concept for the Periodic Table of Software thesis.
Converts handwritten notes and photographs into typeset documents. Claude Vision extracts structure into JSON IR; renders to Typst, LaTeX, or HTML.
Voice-to-paper infrastructure. Record → transcribe → structure with AI → auto-compile to PDF. Zero-cost publishing on free tiers.
Real-time audio sonification of neural network training. Maps layer weights to frequency bands; demonstrates grokking (sudden memorization → generalization phase transition) through sound.
Collection of circuits from 4 to hundreds of qubits. RY, RZ, SX, CZ gate operations executed on quantum hardware. Includes hackathon competition entries.
Daily AI-generated Mad Libs with 16 custom lib types, crowdsourced quality voting, and analytics. Deployed live.
Platformer engine with physics, AI enemy behaviors (chaser, camper, bouncer), level editor, and neural network input vector generation for future ML integration.
TDOA cross-correlation across environmental scenarios (Kathmandu, Amazon rainforest). Temperature-adjusted speed of sound, dual-frequency analysis at 1 kHz and 20 kHz.
Embedded firmware for real-time AC analysis: peak detection, zero-cross detection, min/max tracking, and data averaging at 50/60 Hz. Timer interrupts and ISR callbacks.
Hadamard transforms, Joukowsky airfoil flow fields, Radon transforms, SVD image compression, Fourier phase analysis, and cross-correlation with sliding windows.
Discrete mathematical model simulating wealth redistribution dynamics. Analyzes convergence behavior and equilibrium states in economic systems.
All downloadable reports, research papers, and project documentation.
Specialization in Power & Control. Undergraduate thesis on real-time thermal rating of overhead transmission lines, published at IEEE PowerCon 2024.
