1 of 21

Lessons From The Field: The How of GenAI

Mohamed Shaaban�Global Director of AI Product, SCALE AI

2 of 21

Research & Data

Applications

Pre-Training*

Deployment*

User reported issues

General Post-Training (SFT + RLHF)

Red Teaming

Specialized Post-Training (e.g RLVR)

Model Evaluation

*Pre-Training, Post-training and Deployment are managed by customers.

APPS + AGENTS

DIALECT

DATA + MODELS

SGP AI Infra

AGENT OUTPUTS

SGP Apps and Workflows

DECISIONS

Dialect

EVALS

ENTERPRISE OVERSIGHT LAYER

RED TEAMING

POLICY

INTENT

3 of 21

Everyone’s Freaking OUT!

3

4 of 21

POCs Fail �Because:

The how

The what

5 of 21

The Right What:

  • Digital Context
  • Latent Knowledge
  • Oversight

6 of 21

Digital Context

Data

Models

Integrations

Context

Applications/Tools

01

7 of 21

02�Latent Knowledge

Representation

INDUSTRY EXPERT

Capture tacit → codify into data models, embeddings, training infrastructure

KNOWLEDGE EXTRACTION & STRUCTURING

8 of 21

03�Oversight

CUSTOMIZATION NEEDED

INFRASTRUCTURE

MODEL

SEMANTIC ABSTRACTION

ORCHESTRATION

APPLICATION

9 of 21

The Wrong What:

  • Narrow Vertical Solution
  • Non-Model-Agnostic Bets

10 of 21

POCs Fail �Because:

The how

The what

11 of 21

The Right How:

  • Iterative and Collaborative Forward Deployed Motion

12 of 21

Lessons From Aerospace

  • 1+ Billion Dollars
  • <20 Million Dollars?

13 of 21

Lessons From Aerospace

Systems Engineering

  • Expensive Launch
  • Design everything to work as intended

14 of 21

Lessons From Aerospace

A New Approach

  • Cheap Launch
  • Try things that are not intended for space

15 of 21

Spec Definition

Flight Feedback & Data Analysis

Module & Subsystem Refinement

Launch!

Build Simplest MVP

Closed Loop Engineering

Iterative Engineering Lifecycle

15

15

Confidential | ©2025 Scale Inc.

INTERNATIONAL BASE CAMP

16 of 21

Lessons From Aerospace

  • Systems Engineering
  • Closed Loop Engineering

17 of 21

Lessons From Aerospace

  • Forward Deployed Engineering
  • Software Development LifeCycle

18 of 21

Problem Definition

User Feedback & Data Analysis

Model & Workflow Refinement

Launch! to User

Build Simplest MVP

Forward Deployed Engineering

Iterative Engineering Lifecycle

18

18

Confidential | ©2025 Scale Inc.

INTERNATIONAL BASE CAMP

19 of 21

The Wrong How:

19

Treating AI like normal software

20 of 21

For Successful GenAI POCs:

  • What? Model‑agnostic systems around context, latent knowledge, oversight.
  • How? Forward‑deployed, closed‑loop iterative engineering
  • Result: Durable solutions, compounding results

21 of 21

Thank you!