---

marp: true
theme: default
class: invert
size: 4:3
style: |
img {background-color: transparent!important;}
a:hover, a:active, a:focus {text-decoration: none;}
header a {color: #ffffff !important; font-size: 40px;}
footer {color: #148ec8;}
header: '◇'
footer: 'Simon Chen 2023'

nomnoml:

markedown drawing UML (Unified Modeling Language) diagrams

Snippets

• 1-train.js: generate nomnoml based on work instruction text
• 2-train.js: generate text based on nomnoml input
• 3-openFDA.js: data preparation

---

Proprietary LLM

.. base model + training

• base model: GPT-4 or GPT-3.5 turbo
• training:
  data: secured repository (Azure Cloud)
  alogrithm:
  +vectorstore embedding
  +finetuinging (prompt engineering)
  +nascent tools (eg BLIP2, Salesforce)

---

model #1: data2chart

data: mock wi-320
Tesla Maintenance Manual (https://onedrive.live.com/?cid=597A1F50B291367A&id=597A1F50B291367A%216571&parId=597A1F50B291367A%216234&o=OneUp)

---

training design

• training dataset
  

---

vectorstore: data2chart.index

vlidation_tesla

---

model #2-chart2data

---

model #3-chart 2 codeblock

---

model #4-flows control

---

rich model

-- multiple iterations / multiple epochs

• 
• draft rvision history
• tabluated content
• openFDA

---

openFDA -> auto feed to model

• adverse events
  

---

• FDA API endpoint
  https://api.fda.gov/device/event.json?search=device.generic_name:tomography&limit=1

• Intervalize
  // Poll OpenFDA every 60 minutes setInterval(checkAdverseEvents, 60 * 60 * 1000);

---

Github

https://github.com/metatatt