Nikshep Saravanan

About

Building a consumer product, in stealthToronto

I'm Nikshep, based in Toronto. I explore ideas by building them, at the edges of AI, crypto, and markets.

Infinitely curious, delusional optimist. Right now that's pointed at a consumer product I can't talk about yet. Away from the keyboard: techno, psydub, and mountains with friends.

  • 2exits
  • $50M+protocol volume
  • 790K+search impressions
  • 1,490GitHub stars
  • 1.2Marticle views

Ventures

4 ventures2 exits

I built HumanPlane, a chart room for prediction markets, and Moltlaunch, a work protocol for agents. I exited RealmPlay, and SoulBazaar was acquired before it launched.

Experience

7 roles2016–2026

I was a Kalshi Builder Fellow, and before going independent spent three years as a senior engineer and tech lead at Instacart. Earlier, data infrastructure internships at Coinbase, Deliverr, SeatGeek, and PagerDuty, and computer science at Waterloo.

Projects

22 projects1,490 stars

Some of it finds a real audience: modelgrep and viberank both pull steady search traffic, and homunculus helped inspire the learning system in everything-claude-code.

On the research side, thimble puts structured tool calling into a 48M-parameter model, bankai adapts 1-bit LLMs with tiny XOR patches, and I built the orchestrator behind MC-Bench. Upstream, two of my fixes ship in Nous Research's Hermes Agent. In all: 1,490 GitHub stars across 22 projects.

Starred and followed by 1,521 developers in 50 countries, including people at

GoogleMicrosoftMetaAppleAmazonNVIDIAVercelStripeAirbnbSpaceXTencentByteDanceAtlassian
141 United States32 Germany32 China28 Canada28 United Kingdom20 India14 Japan12 Australia+42 more countries

Also on the list: the creators of everything-claude-code, tokscale, Sablier, Rainbow and egghead.io.

Open source

tools people star and fork
  1. homunculusclaude code · adaptive agents 39847 forks A Claude Code plugin that learns from your prompts and tool use. Turns observed patterns into reusable instincts, skills, and commands.
  2. terminalpolymarket · trading infrastructure 12426 forks Self-hosted Polymarket terminal with live order books, trader history, and wallet-signed orders. Rust backend, keyboard-driven interface.
  3. openvenicemultimodal ai · visual workflows 6413 forks A self-hostable Venice AI frontend for chat, images, audio, music, and video. Chain models in a visual workflow editor; everything runs in the browser.

Search figures from Google Search Console, all data through 11 October 2026; viberank usage from viberank.app/api/stats.

Earlier workStudent projects · 2017–2020

Investing

3 public calls2026

I invest in public, and in writing. I bought Venice around $2, when almost no one was talking about it, and sold near the $34 top. I called NEAR at $1.30 as the settlement layer for agents, and found Nillion while looking for private memory for my own agents.

  • Venice · $VVV · exitedFeb – Sep 2026 17×$2 entry → $34 exit Bought at a $160M cap with almost no one on the timeline talking about it: a private-AI company doing $4M a month, 1.3M users, no VCs. Held through the run, then trimmed and turned short at $34 when the token started pricing in burns that hadn't happened yet.
    First breakdown · ~$3Public call · $11Short at $34Exit thesis
  • NEAR · $NEAR · calledMay 2026 3.3×$1.30 call → $4.30, peak near $5 Wrote it up as the settlement and coordination layer for agents while the market was still comparing its TVL to Solana's. Chain abstraction, intents, and confidential sends were the thesis.
    Thesis · 183K viewsCall at $1.30Recap at $3
  • Nillion · $NIL · holdingMay 2026 – $0.05first call, at a $25M cap Found it while digging for private memory for agents. Sold when its private-AI approach looked weak, then re-entered at $0.04–0.05 the week it pivoted to encrypted markets.
    Inside the Blind ComputerEncrypted markets thesis

Prices and dates are as I posted them at the time; every link goes to the original post. Full track record. Not financial advice.

Writing

31 articles7 essays and reports

I write in public: 31 long-form pieces on X with 1.2M views and 2,400+ bookmarks, from token deep dives to where AI is heading. For the longer arguments, start with Control Surface or Liquid Talent.

Press

6 features

Contact

TorontoEastern time

The best way to reach me is a DM on X, or email.

© 2026 Nikshep Saravanan · TorontoBack to top ↑