Aaryush Gupta

Aaryush Gupta

I build the systems behind how anyone solves problems.

Interesting people are the best part of this. Say hi.

Save contact
Dwg. no.
AG-001
Sheet
01 / 05
Region
us-central
Last deploy
2026.10 · 035b43f
Scale
NTS
§ 02How AI fitsSheet 02 / 05

Four levels where AI changes an organization.

Most problems show up at one level but are caused at another. I start by finding which level a problem actually lives at, then build there. Each level links to the work below that sits at it.

  1. How the organization runs

    Operating models and decision rights: who decides what, and which decisions a system can carry so people don’t have to.

    Seen in CylerityOperating modelGupta AdvisoryAdvisory

  2. How work moves

    The handoffs between people and AI: where judgment is needed, and where volume can be absorbed without anyone touching it.

    Seen in EarlierSmartBenAICylerityEngineering intake

  3. How things get built

    Engineering and product processes rebuilt around AI, so more people can build and fewer steps wait on a specialist.

    Seen in Ready ReboundEngineering loopCylerityEngineering intake

  4. What gets built

    Products that only exist because AI does.

    Seen in EarlierUniviseSkale.winCareer platformCylerityClaims underwritingEarlierUMN Discovery Launchpad
§ 03WorkFig. 3.1 — Timeline, newest first

Companies, and the systems inside them.

Next — contextOS. Coming soon.

Gupta Advisory Founder I work with organizations on how they should operate with AI. Inside Gupta Advisory

Jan 2024 – PresentAI Lead

Cylerity

A fintech-healthcare company that underwrites insurance claims and advances cash to healthcare providers, so they aren't waiting weeks to get paid.

Inside Cylerity
SYS.01 L.04

Claims underwriting

Turns a weeks-long claims cycle into same-day funding decisions. Risk and compliance run in parallel, and providers see what was approved and why.

Backs an eight-figure lending facility
SYS.02 L.02L.03

Engineering intake

Anyone in the company drops an idea into a shared Claude project. A conversation draws out what's needed and classifies it as an idea, tracked, or a scope, ready to build.

Retired a sprawling spreadsheet
2023 – 2024Earlier work

Earlier

The first products I built with AI, from the release of GPT-4 onward.

SYS.05 L.04

Univise

Built a university advising system that spread across campus through word of mouth. No marketing, no push, just a product that solved a daily problem.

Best pitch at Transcend
SYS.06 L.02

SmartBenAI

Automated claims data extraction for healthcare TPAs while keeping humans in the loop where compliance demands it.

Compliance-first automation
SYS.07 L.04

UMN Discovery Launchpad

Replaced ad-hoc startup mentorship with a system that structures what each founder has already done, so the right people can help without relearning it.
§ 04WritingSheet 04 / 05

Notes on where this is going.

Once, for Good

Solving the same problem twice never sat right with me. Here’s what AI changed about that.

This Is Bigger Than Electricity

My attempt, after years of failing to find the right words, to finally explain what AI means for everyone — not the technology, but what's already changing in the world around you, right now, while most people are going about their lives.

The Quiet Singularity

An observational analysis of three concurrent developments in artificial intelligence — recursive model self-improvement, the commoditization of frontier-class inference, and the compounding automation of cognitive labor — and their implications for institutions designed around assumptions of scarcity.
All writing
Skale.win
Esc

Skale.win · Founding member, Product & Strategy · SYS.04

  • L.04 — What gets built
Skale.win career platform system diagram An optional job link goes through Perplexity and comes back as structured job requirements, alongside an optional industry and a difficulty level. An interview planner turns that into a dynamic question plan, reshaped after every answer. The interview itself is a real-time voice loop: speech to text, a language model on OpenRouter, text to speech, back to the candidate in under 500 milliseconds a turn. An evaluator scores competencies plus categories generated for that interview and writes specific feedback. The profile embeds skills and preferences to match roles by vector similarity, and the weakest categories seed the next interview. JOB URL INDUSTRY LEVEL < 500 MSPER TURNSPECIFIC FEEDBACKWEAKEST CATEGORIESSEED THE NEXT INTERVIEW Job context PERPLEXITY · STRUCTUREDN.01 Interview planner DYNAMIC · PER ANSWERN.02 OpenRouter LLMN.03 TTS VOICE OUTN.04 Candidate SPEAKSN.05 STT VOICE INN.06 Evaluator COMPETENCIES + LIVE CATEGORIESN.07 Profile embeddings SKILLS · PREFERENCESN.08 Role matching VECTOR SIMILARITYN.09
Fig. 04 — Every interview sharpens the profile, and the person. FlowThe candidateNext interview, tailored

Hiring runs on resumes: a flat, low-resolution signal of what someone can actually do.

The platform started with an AI interview, a quick five minutes or a full thirty. A job link went through Perplexity and came back as structured requirements; without one, an industry or nothing at all was enough. The interview itself ran as a voice loop, speech to text into a model on OpenRouter and back out as speech, in under 500 milliseconds a turn, and every answer reshaped the next question. Each interview was scored on core competencies plus categories generated for that conversation, and the results became specific, actionable feedback and an embedded profile of the person’s skills and preferences.

That model recommended jobs and aimed the next interview at the weak spots. The more someone used it, the sharper their profile got, and the better they got.

Job data
Perplexity, structured from a job link
Voice loop
Speech to text → OpenRouter → text to speech, under 500ms a turn
Profile
Skills and preference embeddings
Model
White-label & enterprise hiring

Keyboard map

?
This map
g h
Go to the hero
g l
Go to how AI fits
g w
Go to work
g r
Go to writing
g c
Go to contact