Engineer · Researcher · Founder and CEO of Tech4Biz Solutions · Bengaluru

Yasha Khandelwal

I measure what computing actually costs.

Numbers on AI, quantum and edge hardware. The tools, the data, the papers and the weekly write-ups, all published in the open.

0+
years in big systems
0
published studies with DOIs
0
open tools, citable releases
0
IEEE working groups
01

Every era of computing has a currency. Ours is the claim: faster, greener, smarter. I stopped trusting claims and started weighing answers in joules.

A benchmark you cannot re-run is an advertisement. A number without its meter boundary is a story. So everything here ships with the harness, the raw data and the evidence, and holds up when a stranger runs it.

02

The instrument.

An open benchmark that weighs one correct answer, on any machine, on the same scale.

energy-per-result

Vendor neutral, Apache 2.0, monthly releases. Joules per correct result for quantum and classical compute, with every run's evidence kept in the repo.

FIRST TWO BACKENDS · SEPTEMBER 2026

The MacBook beats the datacentre GPU by 40x

0.1 1 10 100 joules per correct result, log Apple M1 0.35 J RTX 4090 14.8 J

A two qubit simulation cannot fill a 450 watt board, so most of its power buys nothing. Fit beats raw power, and now that is measurable.

Inside the pipeline

epr CLI Task registry Apple M1 RTX 4090 QPU Correctness |E − E₀| < 1.6 mHa Aggregator median J / correct Evidence logs · screenshots Zenodo DOI per release one command H₂ VQE · target fixed meter boundary, stated powermetrics NVML, board next backend

Three rules run through it. State where the meter sits. Count only correct answers, so wasted runs make the number worse. Keep raw evidence and mint a DOI, so every release stays citable and checkable forever.

The workload itself

q₀ q₁ RY(θ₁) RY(θ₂) RY(θ₃) RY(θ₄) Classical optimizer E(θ) → update θ entangle measure repeat until |E − E₀| < 1.6 mHa · every shot is metered

The H2 ground state through VQE: a parameterised two qubit circuit, measured, fed to a classical optimizer, looped to chemical accuracy. Millions of shots per correct result, and the meter sees all of them.

03

The evidence.

Published studies and citable tools, each with its own machinery drawn out. Full pages: research and open source.

2026 · ZENODO PREPRINT · SIX-RUN CONTROLLED STUDY

Specification Completeness, Not Model Capability, Determines the Reliability of LLM-Generated Backend Code

The same coding agent built the same backend module six times. The specification, not the model, decided reliability.

Raw spec · 3 runs Build kit · 3 runs Same agent Score reliability tracked per run, all six published

2026 · ZENODO PREPRINT · FIELD STUDY IN PROGRESS

Predicting EV charger degradation from OCPP telemetry

Chargers already stream their own vital signs. evse-drift turns that stream into an early warning.

Charger MeterValues StatusNotification Extractor alert threshold drift score over time

2026 · PYTHON · ZERO DEPENDENCIES

qec-overhead

What a fault tolerant computation really costs: physical qubits and time, as a function of code distance.

data qubits · dots X and Z stabilisers · plaquettes 1 logical qubit ≈ 2d² physical the calculator gives qubits and time per algorithm

2026 · PYTHON

orbital-compute-budget

A data centre in orbit computes on sunlight and survives eclipse on batteries. This tool computes the hours that are actually usable.

Earth eclipse · battery budget sunlit · compute + charge output: effective compute hours per day
04

Writing. Weekly.

Measured numbers, runnable code and full architecture teardowns.

Coming up: quantizing a 3B model and measuring per-layer divergence, what a surface code really costs in physical qubits, and teardowns of production systems I have designed and shipped. Browse the archive

05

The teaching.

Books, video and standards work, because a measurement matters most when others can use it. Full pages: books and talks.

NOW ON AMAZON · KINDLE AND PAPERBACK

AI from First Principles
Volume 1: The Mathematics

Eight chapters that take a reader from what AI actually is to the mathematics underneath it: data, reading formulas without fear, linear algebra, calculus, probability and optimization.

765 pages in premium colour so every one of the 77 diagrams prints clearly, and every derivation is worked and verified.

Buy on Amazon Volume 2, The Machines, in progress

CEO Deep Tech

Deep technology videos: architectures, measurements and how real systems get built.

STANDARDS PARTICIPATION

IEEE P3329 · quantum computing energy efficiency

IEEE P3109 · arithmetic formats for machine learning

IEEE P3419 · LLM evaluation

IEEE P3394 · LLM agent interfaces

Plus talks at Supercomputing India and Women Who Code, and a contribution to the Nasscom AGI Roadmap for India.

06

The company.

Since I took the CEO seat back in 2024, Tech4Biz Solutions has been shipping products of its own alongside client engineering.

SOVEREIGN CLOUD

Cloudtopiaa

A sovereign IaaS platform running on CtrlS Tier-4 data centres, built for workloads that have to stay in-country: government, regulated industry and research.

DETERMINISTIC ENGINEERING

CodeNuk

Deterministic backend generation. The discipline my specification study measures, complete specs before code, turned into a delivery pipeline that produces the same reliable output every run.

Behind the products sits the engineering practice: software, hardware design and applied R&D delivered across the US, EU and India. Explore Tech4Biz Solutions

07

The story so far.

At scale, opinions lose and measurements win. The full story lives on the about page.

BEFORE 2016

Twenty years inside big systems

Principal architect on enterprise systems, with engineering roles at Cisco, Google and SpaceX.

2016

Founded Tech4Biz Solutions

An engineering company in Bengaluru delivering software, hardware design and R&D across the US, EU and India.

2024

Back at the helm

Resumed the active CEO role at Tech4Biz, steering the company deeper into hardware design and applied R&D.

2026

The measurement year

First published studies with DOIs, first book on Amazon, four citable open source tools, four IEEE working groups.

Let's talk.

yasha.khandelwal@tech4biz.io

ORCID GitHub YouTube Wikidata Tech4Biz Solutions