# SimbaStack

> SimbaStack is a US engineering team that builds AI agents and onchain systems. Founded by NJ Singh: 14 years shipping production software, building on Ethereum since 2017, top 1% on Upwork. We ship into the client's codebase by fixed-price sprint or a monthly embedded team; onchain work is scoped separately.

## What we build

**AI agents.** Customer-facing and internal agents wired to real inventory, orders, accounts and money, with hard limits on what they can do without a human. We also take prototypes built with Claude Code, Cursor or Lovable to production: auth, payments, real data and monitoring.

**Onchain systems.** Vaults, token migrations and trading products on Ethereum, Arbitrum, Hyperliquid and Polygon. Clients include Pear Protocol (two years) and QuickSwap (backend engineering).

## Services

- **AI Build Sprint — $12,500.** Two weeks to a working AI feature in your environment. One workflow, agreed acceptance criteria, in your repo, 30 days of support after launch. Bounded scope by design: multi-system builds, complex permissions, or production-critical paths get scoped and priced on the first call. If we can't see a clear first build on that call, we'll say so and won't take the sprint.
- **Embedded team — $10,000/month.** Two engineers plus NJ working your roadmap. Two-month minimum.
- **Onchain projects.** Scoped separately; cash budgets preferred.
- **AI Visibility Audit — $1,500** ([details](https://www.simbastack.com/ai-visibility/)). For software companies. What ChatGPT, Claude, Gemini and Perplexity say about your product (share of mentions across repeated runs, against competitors you pick), what AI crawlers can read on your site, and whether a coding or browser agent can sign up, follow your docs and use it. Five business days, raw answers included, re-test at 60 days. Fixes from $5,000 over two weeks, including an agent-run product blog on SlopIt on your own domain when owned content is the gap. Bigger agent builds (MCP servers, agent sign-up) go to the AI Build Sprint. No ranking guarantees.

## Production work

### Mara Hilltop — AI guest agent for a safari lodge (our own lodge, runs on KaribuKit)

- **Stack:** KaribuKit PMS · custom MCP server · multi-model
- **Problem:** Small team, everything manual. Guest inquiries arriving at 2am. Staff buried in ops instead of guest experience.
- **What we built:** Ranger, a guest-facing agent on the lodge website that answers inquiries around the clock, quotes from live availability in the guest's language, and hands staff a ready-to-confirm booking. A staff-side copilot runs reservations, check-ins, and pricing through chat. On 31 August it booked a $300 game drive for a guest with no one on staff involved.
- **Our role:** Our own lodge, not a client. We built KaribuKit, the property system underneath, and we live with every edge case.
- **Live:** https://marahilltop.com/

### Pear Protocol — Agent Pear, AI trading agent (client)

- **Stack:** Multi-model DAG · on-chain + sentiment pipeline · confirm-before-execute
- **Problem:** Traders had more on-chain data, sentiment, and pair signals than any person can watch, and no way to act on them quickly without giving up control.
- **What we built:** Agent Pear answers pair and market questions from live on-chain stats and sentiment, then stages long/short perpetuals positions that execute only after the trader confirms. Live for Pear's retail traders on Telegram and web.
- **Our role:** Two years as Pear's engineering team, NJ leading. $300M+ has traded through the platform we helped build. We built Pear Vaults end to end (first $1M cap filled in about ten hours, September 2026) and led the PEAR token's move from Arbitrum to Hyperliquid (852M supply, live 19 October 2026). Case study: https://www.simbastack.com/work/pear/
- **Live:** https://pear.garden/

### SlopIt — agent-native publishing platform (ours)

- **Stack:** TypeScript · MCP · REST · Static HTML
- **Problem:** Existing blog tools assume a human at the keyboard — OAuth flows, draft UIs, image uploads. AI agents trying to publish hit a wall of human-shaped friction.
- **What we built:** A publishing platform with the agent as a first-class user. Tell Claude or ChatGPT to publish a post, the agent calls the MCP tool, and a live URL comes back in seconds.
- **Result:** Zero human steps from AI prompt to live URL. Open-core, live at slopit.io.
- **Live:** https://slopit.io

## Team

- **NJ Singh** — Founder & Chief Architect. Penn State. 14 years shipping production software, building on Ethereum since 2017, top 1% on Upwork. Picks up the phone.
- **Deep** — Senior Engineer & Delivery Lead. Leads client delivery end-to-end.
- **Pranjal** — Senior Engineer & Systems Architect (ex-Swiggy).
- **Chetan** — Full-Stack Engineer (Design Lead).
- **Hemanshu** — Full-Stack Engineer (Backend / Performance).

## Notable writing

### While I slept, my 5-year-old MacBook ran Gemma 4 locally and indexed a year of video (May 2026)

A local-LLM pipeline that indexes a year of video on consumer hardware — a 2021 MacBook running Gemma 4 with 50GB swap. Open-sourced as framedex (MIT). Front page of Hacker News for ~30 hours, 470 upvotes, 142 comments.

- **Read:** https://blog.simbastack.com/indexed-a-year-of-video-locally/
- **Discussed on Hacker News:** https://news.ycombinator.com/item?id=48222733

### We rebuilt the structured output problem one layer up (May 2026)

History of the structured-output era and the tool-calling era; argument that the same coercion pattern is repeating one stack-layer up.

- **Read:** https://blog.simbastack.com/we-rebuilt-the-structured-output-problem-one-layer-up/

## Contact

- Email: nj@simbastack.com
- Phone / WhatsApp: +1 (717) 683-9393
- LinkedIn (NJ): https://www.linkedin.com/in/singhniranjan/
- X / Twitter (NJ): https://x.com/bozo_eth
- Upwork (NJ, top 1%): https://www.upwork.com/freelancers/niranjans