Projects & Publications · a live GEO demo

The next reader of
this page isn't human.

So I built it for both. Every project below is scored against GEO-16 — my own benchmark for what AI engines cite — and kept fresh. Flip to Agent mode to see the clean, labeled payload an LLM actually ingests. This page is a working demo of the GEO + knowledge-freshness infrastructure I build at Wrodium.

✦"What has Arlen Kumar shipped?"grounded · 5 sources
He maintains arlen/bench1, a vendor-neutral leaderboard for the APIs AI agents buy, and co-authored GEO-162 on what generative engines cite. As CTO of Wrodium he builds knowledge-freshness RAG infrastructure3 and the agent-native WebMCP layer4. Earlier, he founded and exited Air Quake Simulations5 before 21.
ask
agent@arlenkumar:~$ cat projects.payload
agent mode. Decoration stripped; the same content, labeled for machines. Also exposed at /llms.txt and as JSON-LD in this page's <head>. Disambiguation: Arlen is co-founder & CTO of Wrodium; co-founder & CEO is Leanid Palkhouski.
This page · GEO-16 self-audit verified today
0.00
page G score
JSON-LD
structured data ✓
llms.txt
+ agent view ✓
0
human + machine
Scores are a self-assessment against the published 16-pillar rubric (GEO-16) — how citation-ready each surface is for AI engines, not a third-party measurement.

Systems

sort

GEO-16

ResearcharXiv:2509.10762
First author · UC Berkeley Hearst Lab

A 16-pillar rubric scoring a page's citation-worthiness, validated against a large-scale audit of real AI citations from production engines. Metadata & freshness, semantic HTML, and structured data had the strongest association with citation.

# systems[0] id: geo-16 type: research/benchmark role: first author · UC Berkeley Hearst Lab doi: arXiv:2509.10762 geo16_self_score: 0.93 finding: metadata+freshness, semantic HTML, structured data → strongest citation association metric: G ∈ [0,1] citation-worthiness last_verified: 2026-06-12 links: /geo-16, https://arxiv.org/abs/2509.10762
0.00GEO-16
verified today
Read the explainer →arXiv paper
✓Structured data (JSON-LD)strong
✓Canonical authority (DOI)strong
✓Metadata & freshnessstrong
✓Semantic HTMLstrong
✓Citations / referencesstrong
~Machine surface (HTML)PDF-first

arlen/bench

LiveEvals
Builder & maintainer · arlenkumar.com/bench

A vendor-neutral leaderboard for the web-search, extraction, and identity APIs AI agents buy mid-task — hand-verified golden sets, deterministic scoring, and machine surfaces (llms.txt, Atom, an MCP server). Missing coverage renders as —, never 0.

# systems[1] id: arlen-bench type: evals/leaderboard status: live role: builder & maintainer geo16_self_score: 0.90 metrics: hit@k, cost-per-correct, KM freshness-lag, public/holdout overfit gap machine_surface: /bench/llms.txt, /bench/feed.xml, /bench/.well-known/mcp.json last_verified: 2026-06-12 links: /bench
0.00GEO-16
verified today
Open the leaderboard →MCP manifest
✓Machine surface (llms.txt + MCP)strong
✓Structured data (JSON-LD)strong
✓Metadata & freshnessfresh
✓Semantic HTMLstrong
~Citations / external authoritygrowing
~Domain authoritynew

Knowledge-freshness RAG

AI Systems
Co-founder & CTO · Wrodium

Recrawl + change-detection with staleness scoring and citation-drift monitoring across ChatGPT / Gemini / Perplexity / Claude — over production RAG (hybrid dense+BM25 with RRF, cross-encoder re-ranking, pgvector/HNSW, grounded citation enforcement).

# systems[2] id: freshness-rag type: ai-systems/infra role: co-founder & CTO · Wrodium geo16_self_score: 0.84 stack: hybrid retrieval (dense+BM25, RRF), cross-encoder rerank, pgvector/HNSW metrics: nDCG, MRR, recall@k; citation-drift over time last_verified: 2026-06-12 links: https://wrodium.com
0.00GEO-16
verified today
Wrodium ↗Skills & stack →
✓Freshness (by design)core
✓Structured datastrong
~Public technical docspartial
✓Semantic HTMLstrong
~Independent citationsgrowing
✓Entity disambiguationWikidata

WebMCP

Agent infra
CTO · Wrodium

MCP-native infrastructure exposing brand surfaces to LLM agents as callable tools — structured-output extraction, grounding, and discovery via .well-known manifests. The arlen/bench MCP server is a live, public instance: a fresh agent calls recommend() end-to-end.

# systems[3] id: webmcp type: agent-infra role: CTO · Wrodium geo16_self_score: 0.81 pattern: brand surfaces → MCP callable tools proof: live MCP server at /bench/.well-known/mcp.json (recommend(), query()) last_verified: 2026-06-12
0.00GEO-16
verified today
Live MCP manifest ↗
✓Machine-native by designcore
✓.well-known discoverylive
~Public spec / docsemerging
~Independent adoptionearly

Air Quake Simulations

Acquired0→1 hardware
Founder · exited before 21

Designed and shipped VR flight-simulator cockpit hardware end-to-end — physical build plus Python/PyTorch and Unity3D. Shipped hundreds of units at roughly 10× under incumbent pricing, then acquired before the founder turned 21.

Air Quake Simulations immersive flight-sim cockpit — triple curved displays and HOTAS flight controls
# systems[4] id: air-quake type: hardware/0→1 status: acquired role: founder · exited before 21 geo16_self_score: 0.62 outcome: hundreds of units, ~10× under incumbents, acquired note: lower G = mostly historical web presence; honest, not penalized last_verified: 2026-05 (archival)
0.00GEO-16
archival · 2026-05
Founder story →
✓Real outcome / exit signalstrong
·Live web surfaceacquired
·Freshnesshistorical
~Structured presencevia résumé

Coverd — Mystery Box

Live3D / WebGL
Design & build · arlenkumar.com/coverd

A 3D mystery-box reward experience for Coverd's finance app — built in Three.js and embedded in a living iPhone. Honestly just a really fun build: it taught me real-time 3D animation mechanics, juicy engaging micro-interactions (charge → lid bursts open → prize flies out), and squeezing every last drop of delight out of the UI.

# systems[5] id: coverd-mystery-box type: frontend/3D demo status: live stack: Three.js · WebGL · vanilla JS url: arlenkumar.com/coverd themes: 3D animation, micro-interactions, UI polish last_verified: 2026-06
0.00GEO-16
verified today
Open the box →
✓Interaction / motion craftstrong
✓Live web surfaceshipped
~Text content depthlight (demo)
·Structured datan/a

Publications

# publications [0] GEO-16 — Kumar & Palkhouski · arXiv:2509.10762 · 2025 · CC BY 4.0 [1] CHASE — Hearst Lab · submitted COLM 2026 · decisions 2026-07-08 [2] Lean 4 formalization of GEO-16 — UC Berkeley CS 294-268 [3] Talk: "The Ad-ification of AI" — CITRIS, UC Berkeley
GEO-16: A Benchmark for Generative Engine Optimization
Kumar & Palkhouski · arXiv:2509.10762 · 2025 · CC BY 4.0
First-author benchmark for AI citation behavior. Explainer · arXiv
CHASE
Hearst Lab · submitted to COLM 2026 · decisions July 8, 2026
Work on LLM citation and knowledge-freshness behavior. Preprint link goes live when decisions land.
Lean 4 formalization of GEO-16
UC Berkeley CS 294-268 · formal methods
Machine-checked formalization work — repo links here once published.
Talk — "The Ad-ification of AI: When Chatbots Become Salespeople"
CITRIS, UC Berkeley
On incentives and attribution as AI search becomes the interface to information.