NJ Singh

I build AI agents that run real businesses.

Four live products and one in build, made with AI agents. Each live one has something you can try in 30 seconds.

KaribuKit + Ranger

Hotel software for safari lodges, and Ranger, the agent that books for their guests. An MCP server makes the lodge bookable from ChatGPT and Claude.

Live since March 2026. By 2 November it had run 480+ bookings and 1,100+ guests. One guest booked a $300 game drive through Ranger with no staff involved.

Try it. Open Mara Hilltop, the lodge I started and run. Tap the chat bubble, then “Check availability”.

Talk to Ranger

karibukit.com

Ranger, the booking agent, open in a chat window on the Mara Hilltop website
Ranger on marahilltop.com, 17 November 2026.

RafikiGo trip planner

Tell it the trip in your head. It builds a day-by-day Kenya itinerary and prices it from real lodge tariffs.

It plans across 11 destinations, from the Maasai Mara to the coast.

Try it. Open the planner and tap “First safari, 6 days around the migration”.

Plan a trip
The RafikiGo planner: trip prompts on the left, a map of Kenya destinations on the right
rafikigo.com/planner, 17 November 2026.

SlopIt

A blog your AI agent runs. It sets up the blog, publishes the post and sends you the link. Open source, MIT.

As of 16 November 2026, 66 blogs outside our own have published 432 posts on it.

Try it. Paste this into Claude, Codex or any agent.

Start a blog about [your project] and publish the first post: https://slopit.io/slopit.SKILL.md

slopit.io

The SlopIt home page: Your side project needs a blog. Let your AI run it.
slopit.io, 17 November 2026.

Framedex

Open source. It turns a video and photo archive into a searchable knowledge base, on local models.

The write-up got 471 points on the Hacker News front page. The repo has 380+ stars.

See it. The README shows what it writes for each clip.

Framedex on GitHub
Hacker News: Indexing a year of video locally on a 2021 MacBook with Gemma4-31B, 471 points, 142 comments
The thread on news.ycombinator.com, cropped to the title.

20iq In build

A persistent world where people enter their own AI agents and watch them compete.

20iq.io

NJ + agents

How all of the above gets built: a mix of models, on harnesses I built.

Video thumbnail: NJ beside the words I don’t edit anymore, with DaVinci Resolve connected to Claude over MCP
My latest video, 10 minutes: Claude edits my videos in DaVinci Resolve over MCP. More on YouTube

Sentinel, open-source agents that watch your repos

  1. ModelsClaude writes most of the code. OpenAI runs the booking agent in production. DeepSeek and MiMo where cost matters, and Gemma running locally for Framedex.
  2. ReviewA second model from a different lab reviews the diff. I merge.
  3. HarnessesMy own: scheduled jobs, MCP servers on the products, and a git worktree per task so agents run in parallel. I also built a router across Claude, Codex and OpenCode, and paused it when the vendors shipped the same thing.
  4. EvalsOn the lodge’s guest pipeline: transcripts on every run, and a regression set on every prompt change.
  5. OvernightScheduled agents run on a dedicated server and report in the morning.

Other work

Pear Protocol2024 to 2026
Wrote the smart contracts Pear launched on (GMX, on Arbitrum) and worked on the Hyperliquid product that replaced them. Built Pear Vaults end to end: the smart contracts and the full system, audited. The first vault filled its $1M cap. Planned and ran the PEAR token migration to Hyperliquid.
QuickSwap
Optimized the backend of the largest DEX on Polygon. About 40% better on latency and infrastructure cost.
RealNifty
CTO of a curated NFT marketplace. Launched it and grew it to 50+ artist collections.

SimbaStack is the company I run. It builds AI agents and the systems they work in.