Back to Projects
Mozisha
AI/ML

Mozisha

Full-Stack & AI Engineer

Next.jsNestJSPostgreSQLDrizzlepgvectorMCP SDKOpenAI

About the Project

Mozisha (Taste Engine) is infrastructure for human-led AI orchestration: it captures the moments a person rejects an AI output and turns them into reusable constraints, so hard-won judgment accumulates instead of evaporating. The pieces are deliberately asymmetric — capture has to be nearly free, so it lives inside the editor as an MCP server rather than as a web form, while review, encoding and sharing happen in a full dashboard. Alongside the product there are two standalone clickthrough demos built for funder presentations, which compress a founder's multi-month journey into a guided walkthrough with no backend behind it.

What It's Made Of

MCP capture server

MCP SDK · stdio transport

The capture layer, exposed as a Model Context Protocol server with essentially one tool — log_rejection — so recording why an AI output was rejected costs a single call from inside the editor where the work is happening. Running as a separate transport off the same backend codebase means the tool writes through the same services and permissions as the dashboard, with no second implementation to drift.

Core API

NestJS 11 · PostgreSQL · Drizzle · pgvector

Business logic, auth and permissions, the Judgment Loop state machine, and vector search. Constraints are embedded and stored in pgvector so they surface by conceptual likeness rather than keyword overlap — what matters later is finding the constraint that applies when a situation rhymes, not when it repeats. It also generates portfolio documents directly from validated judgment data.

Dashboard

Next.js 16 · React 19 · Supabase auth · Recharts

Where a rejection becomes a constraint: reviewing captured rejections, walking the six-step Judgment Loop that encodes one into a reusable rule, and browsing the constraint library with personal and organisational views. It also hosts the portfolio generator, which assembles evidence of expertise from the judgment record, and the shared Constraint Commons view.

Funder demo clickthroughs

Next.js · Framer Motion · hardcoded content

Two standalone presentation apps that walk an audience through the methodology — a founder's journey compressed into a guided, animated walkthrough with every value hardcoded and no backend attached. Built to make the method visible and credible in a room, which is a different job from the product and correctly a different codebase.

Key Highlights

  • Built the capture layer as an MCP server exposing one tool — log_rejection — so a rejection is recorded where the work happens rather than in a separate app nobody opens
  • Ran the MCP server off the same backend as the dashboard, so capture writes through the same services and permissions instead of a second implementation that drifts
  • Designed the Judgment Loop, a six-step workflow that turns an individual rejection into an encoded, reusable quality constraint
  • Implemented vector similarity search over pgvector so constraints surface by conceptual likeness rather than keyword overlap
  • Built a portfolio generator that assembles evidence of expertise directly from validated judgment data
  • Designed the Pan-African Constraint Commons for sharing constraints across an organisation and beyond it
  • Shipped two backend-free clickthrough demos for funder presentations, compressing a multi-month founder journey into a guided walkthrough

Technical Challenges

The product only works if capture is nearly free. Anyone will agree that recording why they rejected an AI output is valuable, and nobody will switch to another tool to do it — which is why the capture layer is an MCP server with a single tool rather than a web form, so the cost of logging a rejection is one call from inside the editor. The modelling problem underneath is harder: 'taste' is tacit, and a constraint is only reusable if it generalises past the specific output that triggered it. Vector search over embeddings is what makes a stored constraint findable later, when the situation rhymes rather than repeats.