Mazaal AI

Mazaal is an AI automation platform where users compose workflows out of agents and pieces. An agent is an LLM-driven decision step; a piece is a deterministic action (HTTP call, transform, branch, file write, integration with third-party services). The runtime executes these in a controllable graph with retries, timeouts, and credit accounting.

Why this split

Pure-LLM workflows are unpredictable. Pure-deterministic workflows are not AI. The interesting product surface is somewhere in the middle — and that means you need first-class primitives for both, with a clear handoff.

The credit system has two axes: task credits (per workflow run) and AI credits (per LLM call, with different costs per model). Users buy credits; the platform debits credits atomically per step. Failed steps refund automatically. This is the kind of plumbing that nobody talks about until it is wrong — and then it is the only thing that matters.

Stripe + QPay

International payments through Stripe. Domestic Mongolian payments through QPay. Both flow into the same ledger, with reconciliation reports that match payouts to platform credit grants.

The provider key / platform key model is what makes the platform extensible. A third-party developer can bring their own LLM provider key, or use platform credits. They can publish pieces that other users drop into their workflows. ActivePieces integration means we sit on top of an existing open ecosystem of connectors instead of trying to build everything ourselves.

AppSumo launch

Launched on AppSumo as a distribution channel — a real test of whether the product could survive non-technical buyers self-onboarding.

I would invest in usage analytics and credit-forecasting tooling earlier. The hardest product question at scale is "what will this user's bill be next month," and the answer requires instrumentation that is easy to defer.

I would also separate agent runtime from platform billing sooner. Mixing them in a single deployable slowed down both.