Builder · Founder · Researcher

Arlen
Kumar

CTO @ Wrodium · NLP/IR @ Berkeley Hearst Lab · GEO-16

↓

I build products, platforms, GEO infrastructure, AI benchmarks, agent-readable surfaces, and the systems that fight bureaucracy and measure AI.

  • products.
  • platforms.
  • GEO infra.
  • benchmarks.
  • RAG systems.
  • agent layers.
  • web apps.
  • APIs.
  • backends.
  • pipelines.
  • dev tooling.
  • living articles.
  • cockpits.
  • foundations.

CTO at Wrodium. NLP/IR researcher at Berkeley's Hearst Lab. Currently building the knowledge-freshness infrastructure that decides what AI engines cite — and writing autopsies for dead benchmarks.

Read the research → See what I'm building

Berkeley · SkyDeck · Hearst Lab

UC BerkeleySkyDeck '24Prof. Marti HearstAir Quake (exit)SCET featured
UC BerkeleyCS, Data Science, Economics
Hearst LabResearch under Prof. Marti Hearst
Berkeley SkyDeck '24Backed · Most Innovative Tech
Previous exitAir Quake Simulations (acquired)
● Now — June 2026

What I'm actually working on.

The honest version. Three things, each with a real deadline.

CHASE @ COLM 2026

Research paper submitted to COLM 2026 — the Conference on Language Modeling. Awaiting the decision round.

Under reviewdecisions Jul 8, 2026

Wrodium

Building the GEO + knowledge-freshness infrastructure that decides which sources AI engines trust and cite.

$1.5M pre-seed2026
Things I build

Shipped, live, and clickable.

All projects & research →
Selected work

Things that exist and you can check.

Every card links to something live.

Open dataset · Evals

The Benchmark Graveyard

Autopsies of dead AI benchmarks — saturation, contamination, gaming — as a citable dataset with reproducible saturation curves and evidence tiers.

Visit the museum →Code & data ↗
Formal methods

proofwars

Competitive Lean 4 theorem proving — a full Glicko-2 rating engine, a sandboxed warm prover pool, and an authoritative match server.

Play a duel →Code ↗
Applied ML

regress.fish

An honest fishing forecast on real NOAA data with a public, auto-published accuracy scoreboard — it ships the Brier score even when it's embarrassing.

Open the forecast →Code ↗
Wrodium, in action

From one prompt to a publishing pipeline.

Describe a content workflow in plain English; the Copilot builds the agent pipeline — ground, generate, de-slop, publish.

External demo

See Wrodium compose a pipeline, live

Loads an interactive app from wrodium-dashboard-ui.vercel.app — nothing third-party loads until you click.

Live demo · the Workflow Copilot composing a daily-blog → CMS pipeline from one prompt

The mission

Hallucinations aren't just embarrassing — they're expensive. I build systems that give LLMs perfect context: structured, verified, current.

Most American business content is a mess — outdated pages, unstructured data, no fact-verification layer. Wrodium is the infrastructure layer that makes AI accurate about real businesses. The goal isn't just to be cited — it's to be cited correctly.

Leanid & Arlen, co-founders of Wrodium
Leanid & Arlen — co-founders of Wrodium
About me

The short version.

Arlen Kumar

At 19 I built and sold Air Quake Simulations, a VR flight-sim hardware company whose 3D-printed cockpits undercut the industry roughly 10×.

Today I research how AI answer engines decide which sources to cite at UC Berkeley's Hearst Lab under Prof. Marti Hearst — the work that became the GEO-16 framework.

And I'm CTO of Wrodium, building the infrastructure that decides who gets cited when AI becomes the interface to everything.

What we're building

Wrodium — Generative Engine Optimization for brands that refuse to be invisible.

Leanid and I started it because the companies that structure their content for AI today will own their categories tomorrow. Everyone else will wonder where their traffic went.

Audits

We show you exactly how AI systems currently describe your brand. (Spoiler: it's probably wrong or incomplete.)

Optimizes

We restructure your content so AI engines can parse, verify, and cite it — schema markup, semantic structure, fact-verification layers.

Maintains

AI changes fast. We keep your content current so you don't drift into hallucination territory.

Measures

We track AI visibility and citations so you can tie "ChatGPT mentioned us" to actual revenue.

We've built integrations for six major CMS platforms. We run experiments constantly. We know what works because we test it — not because we read a blog post about prompt engineering.

Research

Most "AI SEO" advice is vibes. I wanted data.

Leanid and I ran hundreds of experiments on how AI answer engines select sources. The result: "AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO-16 Framework" — a paper identifying the 16 factors that predict whether AI cites you.

It's not theoretical. It's built from real queries, real responses, and real citation patterns across ChatGPT, Gemini, Perplexity, and Claude.

I treat research as product development: every insight becomes a feature, every experiment becomes something I can ship.

Read GEO-16 on arXiv →
GEO-16 research framework
What I care about

The principles under the work.

Honest attribution

The web has a credit problem — creators of accurate, useful content get erased when AI summarizes them with no link. Better structure, better standards, better tracking for how AI cites sources.

Responsibility that ships

I've studied how AI systems erase or misrepresent people — that's a product concern, not just an ethical one. I build systems that optimize for accurate exposure, not just exposure.

Theory → buttons

I like research. I like shipping more. The best part of Wrodium is turning abstract ideas like "generative engine optimization" into dashboards, workflows, and reports people actually use.

Track record

What I've started, sold, and shut down.

Air Quake Simulations immersive flight-sim cockpit — triple curved displays and HOTAS controls

Air Quake Simulations

Acquired

A VR flight-simulator hardware company. Built affordable 3D-printed cockpit components that undercut the industry by 10×.

Read the story →

Wrodium

Current

Co-founded with Leanid Palkhouski. Berkeley SkyDeck-backed. Building the GEO platform for AI-era discovery.

UC Berkeley Research

Ongoing

Studying AI citation behavior under Prof. Marti Hearst. Developed the GEO-16 predictive framework.

A red Australian labradoodle puppy from Catherine Manor Labradoodles

Catherine Manor Labradoodles

Shut down

A community network for Silicon Valley dog owners. Shut down during COVID. (Yes, really.)

Read the story →
Speaking

GEO Conference 2025.

Speaking at GEO Conference 2025

"Why Your Content Isn't Getting Picked by AI"

Leanid Palkhouski, Co-founder @ Wrodium

San Francisco · December 15, 2025

Building at the edge of AI, law, or evals?

I like comparing notes with people making real systems — researchers, founders, and anyone fighting bureaucracy with code.

Let's talk →