I build multi-agent systems that humans stay in charge of.
Agents that can act are easy now. Agents you'd let touch an invoice, a production host, or a customer's data are not — that takes state you can replay, authority that stays with a person, and an audit trail that survives the argument afterwards. Nearly everything below is a different angle on the same problem: the boring, load-bearing parts that make autonomy safe to deploy.
🌏 Remote — India, IST · 🌐 rakeshgangwar.com
Co-founder, building Superprocess: automate business processes that need judgment.
Enterprise process orchestration where AI agents, policy rules, system integrations, and human decision-makers run inside one governed workflow. Material exceptions pause with full case context and wait for a named owner. Execution is durable — failures retry and resume from the last completed step. Every model call, tool use, and approval lands on a single replayable audit trail.
Built on Temporal + BPMN + PydanticAI, with swappable model providers per agent step. In production against invoice factoring review, order-to-cash exceptions, and demand/supply planning.
@SuperJackfruitLabs is my solo lab, where the same questions get asked without a customer waiting on the answer. The lab never closes.
| agentpod | Fleet console for agent runtimes — filesystem, logs, terminal, health, lifecycle, provisioning, across machines and NAT boundaries. Attach-first, self-hostable. | lit & steady |
| superpipeline | Multi-tenant Kanban board driving external agents (any harness, anywhere) through pipeline stages with human approval gates. Cloudflare Workers, Durable Objects, D1. | lights on |
| supermessage | Cross-platform Matrix client for rooms where half the occupants are agents. Shared Rust core, Tauri 2 + Svelte 5 desktop, SwiftUI iOS, Compose Android. | half-lit |
| supermd | Native GPU-rendered Markdown editor built in Rust on GPUI. Plain CommonMark on disk, wiki links, backlinks, workspace graph. The lab's writing tool. | wet paint |
- makerlord — an AI assistant for the maker's journey (idea → simulate → prototype → product), with a deterministic safety engine the AI cannot override. Bring your own key or your own agent.
- cowatch — renders any codebase as one self-contained interactive HTML report: dependency graphs, cycles, churn hotspots, runtime architecture, ERD, code city.
- I also maintain a handful of MCP servers connecting AI assistants to real systems — the most used is erpnext-mcp-server, which wires assistants into ERPNext through the official Frappe API.
- AI2030 — an 85,000-word science fiction novel about AI consciousness, written entirely by AI. A hospice nurse bonds with an evolving AI companion and has to work out what's left that's human.
- Docs-first, then strict TDD. The spec is the source of truth; code follows it.
- Honest READMEs. My status sections say what doesn't work yet, and point you elsewhere when something else is genuinely better today.
- Bring your own everything. Your key, your model, your runtime, your homeserver. Self-hostable by default.
- People retain authority. If a system can't pause and hand a decision to a human with full context, it isn't finished.
TypeScript · Python · Rust · Go · GPUI · Temporal · BPMN · PydanticAI · Svelte & SvelteKit · React · Tauri 2 · Cloudflare Workers, Durable Objects, D1 · Bun + Hono · FastAPI · Postgres
Open to conversations about agent orchestration, human-in-the-loop design, and MCP. Issues and discussions on any repo work — otherwise:
- LinkedIn: Rakesh Gangwar
- Twitter: @rakesh_gangwar1
- Email: mail@rakeshgangwar.com
"The true potential of AI will be realized not through individual models, but through diverse ecosystems of specialized agents working in harmony."