HomeJemari Sapp

AI systems & automation

I design and build AI systems people actually use.

I’m Jemari Sapp, based in Atlanta, with an MS in Information Systems focused on AI for Business Innovation from Georgia State. I turn slow, manual work into systems where AI does the tedious part and a person approves what matters. I design each system, direct AI coding tools to build it, and run it in production.

View resumejemari@jemarisapp.comLinkedInGitHub

How I work

  1. 01Find the tedious part

    Start from the work people already do by hand, and what it costs them each week.

  2. 02Design the system

    Data model, triggers, prompts and the points where a person has to approve.

  3. 03Build with AI

    Direct Claude, Codex and ChatGPT as the development team, reviewing what they produce.

  4. 04Keep facts out of the model

    Prices, policies and availability come from code and real data. The model handles language.

Systems

marks where a person approves before anything moves on.

Kampgrounds of America · AI engineering internship

Multi-agent group sales automation

A multi-agent pipeline that finds group sales leads, qualifies and scores them, checks availability and drafts outreach for a sales manager to approve.

The problem. Each lead took hours of manual searching, routing, availability checks and email drafting before anyone could act on it.

  1. Discover events
  2. Qualify
  3. Check availability
  4. Score
  5. Draft outreach
  6. Manager approves
My role
Built the review dashboard and pipeline monitoring, connected it to Supabase, and moved agent execution behind server routes to protect API keys.
Results
  • A manual workflow cut from hours to under 10 minutes, with final review kept in human hands.
  • Airia agents
  • Supabase
  • React
  • TypeScript
  • Server-side API routes
Read the case study →

Team project · led AI systems design

Navigate IDD

Turns Medicaid waiver policy into multilingual explainer videos, with every script grounded in the official documents and approved by an expert.

The problem. Families need clear answers about disability services, but the policy is dense, mostly in English, and videos were slow to make and quick to go out of date.

  1. Embed policy in Pinecone
  2. Draft from retrieved sections
  3. Expert approves
  4. Translate
  5. Generate avatar video
  6. Approve and publish
My role
Designed the system: retrieval rules, prompts, the approval gates and the spreadsheet control panel across seven connected workflows.
Results
  • Content creation time cut by about 90%.
  • Videos in five languages.
  • Every script traceable to verified policy text.
  • n8n
  • OpenAI
  • Claude
  • Pinecone
  • Google Sheets as control panel
  • Synthesia
  • Blotato
Explore the pipeline →
Navigate IDD explainer video

My product · in production

Trilo

A Discord bot that runs College Football and Madden dynasty leagues: it reads schedule screenshots with AI vision, creates every matchup channel and tracks results.

The problem. Volunteer commissioners spend hours every week on setup and tracking, and leagues fold when they burn out.

  1. Screenshot uploaded
  2. AI vision reads matchups
  3. Commissioner confirms
  4. Channels and players tagged
  5. Results tracked by reaction
My role
Founder. Designed the product, pricing and licensing, led a normalized v2 data model, and built it with AI coding tools.
Results
  • 11 leagues and 133 players.
  • 2,152 commands run since January.
  • 97% of commands complete without an error.
  • Python
  • discord.py
  • OpenAI vision
  • Supabase / Postgres
  • Stripe
  • Railway
See the product →
Trilo app concept screens

My studio · in production

TRISP session system

The system my photography studio runs on: every client session moves through six stages with automatic emails, and ChatGPT sets new sessions up from my calendar.

The problem. Every session meant repeated follow-ups about status, locations, galleries, favorites and delivery dates.

  1. Ask ChatGPT to set up a session
  2. Finds it on Google Calendar
  3. Creates the Supabase record
  4. Client portal link sent
  5. Emails at each stage
My role
Designed the workflow, data model and interfaces, and built it with Codex as my development team.
Results
  • Every session since April has run through it.
  • About an hour of admin saved per client, by my estimate.
  • React
  • Supabase
  • Vercel serverless
  • Resend
  • ChatGPT with Supabase and Calendar connectors
  • Codex
Read the case study →
Frames from the session system explainer video

More systems

Retired

Nova

An AI booking assistant that answered questions on trispstudio.com and offered real open times from my calendar.

  1. Visitor asks
  2. Pinecone knowledge base
  3. Live calendar slots
  4. Slot guard checks the reply
  5. Logged to Notion, email alert

Started on Langflow, then rebuilt natively. A guard compares every reply against real open slots and rewrites any time that isn’t actually free.

  • OpenAI
  • Pinecone
  • Google Calendar
  • Notion
  • Resend
  • Vercel serverless

TRISP Studio

Email booking agent

Reads photography inquiries as they arrive and drafts a reply with real availability, ready for one-click approval.

  1. Gmail push event
  2. Claim thread with a label
  3. Gemini extracts the request
  4. Calendar and pricing from code
  5. Draft for approval

Prices and policies live in code, not in the prompt, so the model handles tone while facts stay fixed. Gmail labels act as the state machine, so no database is needed.

  • Replies drafted in under 30 seconds instead of 15 to 20 minutes.
  • About $27 a month to run, down from about $70 on a no-code platform.
  • TypeScript
  • Vercel serverless
  • Gmail API
  • Google Pub/Sub
  • Gemini

Legal operations

AI-assisted contract generation

Turns structured client intake into jurisdiction-appropriate draft contracts for an attorney to review and approve.

  1. Normalize intake
  2. Assemble prompts and templates
  3. Generate first draft
  4. Attorney reviews

Designed the workflow, prompts and integration, with attorney control over every final document.

  • LLMs
  • Structured outputs
  • Prompt engineering
  • Workflow automation
Read the case study →

TRISP Studio

Code-based video toolkit

Explainers, promos and re-cuts written as code: scenes are web pages rendered frame by frame, timed to narration word by word, with music synthesized in code.

  1. Brief and reference
  2. Style guide and shot list
  3. Approve the plan
  4. Render as code
  5. Review frames

Designed the system and its approval gates, and packaged it as five reusable Claude skills.

  • Python
  • Playwright
  • FFmpeg
  • faster-whisper
  • Claude skills
See how I direct AI →

Stack

AI

  • OpenAI
  • Claude
  • Gemini
  • RAG with Pinecone
  • Multi-agent workflows
  • Structured outputs
  • n8n
  • Airia
  • Langflow

Backend & data

  • Supabase / Postgres
  • Data modeling
  • Vercel serverless
  • Railway
  • Python
  • TypeScript

Integrations

  • Google Workspace APIs
  • Gmail and Pub/Sub
  • Stripe
  • Discord
  • Notion
  • Resend
  • Synthesia

Product & front end

  • React
  • Next.js
  • Product design
  • Design systems
  • Figma

Education

MS, Information SystemsAI for Business Innovation · Georgia State University, 2026
BFA, Graphic DesignGeorgia State University, 2023

I’m also a designer and art director. If you want to see that side, the design portfolio is here.