Run it manually
Role-play how the business receives, evaluates, records, and follows up with one lead today.
Day 1 presentation1 / 31
Zero to MVP AI Bootcamp
From real life to software & AI
Start with a familiar business workflow, then see how software and modern AI turn it into a scalable system.
Today you ship
A practical mental model of software and AI, plus a working project shell for the AI Lead Assistant.
Scroll to move through the presentation.
Learning objectives2 / 31
By the end of today
- 1Translate a familiar business workflow into software components.
- 2Explain frontend, backend, API, database, authentication, and cloud using everyday analogies.
- 3Compare how a human brain learns and remembers with how an AI model is trained and prompted, without treating them as equivalent.
- 4Trace the major steps from rule-based AI to machine learning, deep learning, transformers, and LLMs.
- 5Distinguish an LLM, RAG system, tool-using workflow, and agent.
- 6Choose a model category and provider using quality, speed, cost, privacy, modality, and control.
- 7Apply the complete mental model to the shared AI Lead Assistant and start its project shell safely.
Run of show3 / 31
Today’s learning path
- Module 120 min
From business workflow to software
- Module 230 min
How software works
- Module 320 min
Your brain and an AI model
- Module 425 min
From early AI to LLMs
- Module 525 min
LLMs, RAG, tools, and agents
- Module 630 min
Choose the right AI model
- Guided build
AI Lead Assistant · Stage 1
Build, test, checkpoint, and ship.
Module 1 · Learn · 20 min4 / 31
From business workflow to software
Software starts with an operating model people already understand. Someone receives information, applies business rules, makes a decision, records the outcome, and communicates the next action. Software makes that workflow consistent and scalable by assigning each responsibility to a digital component.
A real example
Without a connected system, a relationship manager interviews a prospect, records answers in a spreadsheet, asks a specialist by email, updates the CRM, and sends separate follow-ups. Our product connects that same operating workflow.
Technical terms
- Business rule
- An explicit condition that tells software what decision or action should follow.
- Human-in-the-loop
- A design where software handles routine work while a person reviews sensitive, uncertain, or exceptional cases.
Module 1 · Reference5 / 31
The same job in two worlds
| Concept | People-led workflow | Digital capability |
|---|---|---|
| Receive information | Meeting, email, or document | A digital form or conversation |
| Pass the request | Email or internal handoff | A secure system connection |
| Apply rules | Coordinator or specialist | Automated decision rules |
| Keep a record | Spreadsheet or CRM entry | Searchable shared storage |
| Recognize patterns | Human experience | Pattern-based assistance |
| Follow up | Manual email or call | Triggered messages and tasks |
| Handle exceptions | Manager judgment | Escalation to a person |
Module 1 · Apply it6 / 31
Apply the concept
Map a familiar business workflow before translating it into software.
- 1Choose a workflow you know well, such as customer qualification, procurement approval, or employee onboarding.
- 2Mark its trigger, intended outcome, and the people involved.
- 3Identify the information, decision, record, handoff, and exception at each stage.
- 4Mark where human judgment and accountability must remain.
Success looks like
A one-page business workflow map that Module 2 can relabel with the correct software components.
Watch for
- Describe the current workflow, not an ideal future solution.
- Keep technology names out of the map until Module 2 introduces them.
Module 2 · Learn · 30 min7 / 31
How software works
A software product is a conversation between specialized parts. The frontend collects an action, an API carries a structured request to the backend, the backend applies protected rules, and the database stores or returns information before a response travels back.
A real example
Think of a restaurant. The menu and dining room are the frontend, the waiter carries a precise request like an API, the kitchen is the backend, the order book is the database, the manager verifies staff identity, and the building is the cloud.
Technical terms
- Frontend
- The visible interface running on the visitor's device.
- Backend
- Protected code that applies business rules and connects data or services.
- API
- An agreed way for software systems to request information or actions from each other.
- Database
- Structured, durable storage that software can create, search, update, and protect.
Module 2 · Reference8 / 31
Software parts in plain English
| Concept | Job | Restaurant analogy | Lead Assistant example |
|---|---|---|---|
| Frontend | Serve the visible experience | Menu and dining room | Questions and result screen |
| API | Carry an agreed request/response | Waiter's order ticket | Submit lead answers |
| Backend | Apply protected rules | Kitchen | Validate and qualify lead |
| Database | Store durable records | Order book | Leads and recommendations |
| Authentication | Verify identity | Staff badge | Business owner sign-in |
| Cloud | Run the system online | Restaurant building and utilities | Public app and server |
Module 2 · Apply it9 / 31
Apply the concept
Trace a submitted form through frontend, API, backend, database, and back to the result screen.
- 1Start with the visitor pressing Submit on the frontend.
- 2Write the API request as an envelope containing an address, action, and data.
- 3Let the backend validate identity and input before reading or writing the database.
- 4Return a response and update the frontend for success, validation failure, or server failure.
Use this prompt
Explain what happens after a visitor submits this form. Trace frontend → internet → API → backend → database → response → frontend. Include authentication, success, validation failure, and server failure. Use a restaurant analogy and then the correct technical terms.
Success looks like
You can trace one click in both directions and explain why private keys and protected rules belong in the backend.
Watch for
- An API is not the backend itself; it is the agreed doorway used to request backend work.
- A database stores information, while a backend decides what may be done with it.
Module 3 · Learn · 20 min10 / 31
Your brain and an AI model
The brain is a useful starting analogy, not a literal blueprint for AI. Both receive input, detect patterns, use recent context, and produce a response. A human also has a body, lived experience, goals, emotions, values, and genuine understanding that a language model does not possess.
A real example
A child learns the idea of a dog through experiences in the world. A language model learns statistical relationships among tokens from large datasets, then predicts a likely next token when prompted.
Technical terms
- Training
- Adjusting a model's internal numerical weights by learning patterns from many examples.
- Inference
- Using a trained model to produce an answer for a new input.
Module 3 · Reference11 / 31
Useful analogy, important differences
| Concept | Human | AI model | Important limit |
|---|---|---|---|
| Input | Senses and lived experience | Tokens, images, audio, or other data | The model receives only encoded data |
| Learning | Experience, teaching, reflection | Training adjusts numerical weights | Training is not human understanding |
| Current focus | Attention and working memory | Prompt and context window | Context is limited and temporary |
| Memory | Personal and semantic memory | Weights plus external storage | A model does not remember a life |
| Response | Intentional thought and action | Predicted output tokens or actions | Plausible output may be wrong |
| Responsibility | Can hold values and accountability | Follows patterns and instructions | Humans remain accountable |
Module 3 · Apply it12 / 31
Apply the concept
Build a brain-versus-AI map and mark where the analogy stops being accurate.
- 1Match senses to model input, attention to relevant context, long-term learning to trained weights, and an action to model output.
- 2Write where each comparison is useful.
- 3Mark the limits: consciousness, lived experience, emotion, values, intent, and accountability.
- 4Explain why a fluent answer can still be false and must be verified.
Use this prompt
Create a careful brain-versus-LLM comparison for a non-technical corporate audience. Separate useful analogies from important differences. Do not claim that an LLM thinks, understands, remembers, or feels exactly like a human.
Success looks like
You can use the analogy to explain AI while clearly stating where it breaks down.
Watch for
- Do not describe model parameters as tiny facts or neurons as exact copies of biological neurons.
- Fluency is evidence that a model predicts language well, not proof of truth or understanding.
Module 4 · Learn · 25 min13 / 31
From early AI to LLMs
AI did not begin with chatbots. The field moved from hand-written rules, through systems that learned statistical patterns, to deep neural networks and transformers trained on enormous datasets. Each wave expanded what machines could do, but none removed the need for clear goals and verification.
A real example
An early spam filter might use rules such as 'if the subject contains FREE, flag it.' A machine-learning filter learns patterns from labeled email. A modern language model can explain why a message looks suspicious, but may still make a confident mistake.
Technical terms
- Transformer
- A neural-network architecture that learns relationships across sequences using attention.
- Foundation model
- A large pretrained model that can be adapted to many downstream tasks.
Module 4 · Timeline14 / 31
How we arrived here
From rules to generative AI
1950s–1960s
The field takes shape
Researchers ask whether machines can reason and create early symbolic programs.
1970s–1980s
Expert systems
Humans encode specialist knowledge as rules such as IF condition THEN action.
1990s–2000s
Statistical machine learning
Systems learn patterns from labeled data for search, spam, speech, and recommendations.
2010s
Deep learning
Larger neural networks, datasets, and compute improve vision, speech, and language.
2017
Transformer architecture
Attention mechanisms enable models to learn relationships across long token sequences efficiently.
2018–2021
Foundation models
Large pretrained models adapt to many tasks through prompts or additional training.
2022–today
Generative and agentic products
Chat, multimodal input, retrieval, tools, and controlled action loops become product building blocks.
Module 4 · Reference15 / 31
The AI family tree
| Concept | Plain-English meaning | Familiar example |
|---|---|---|
| Artificial intelligence | The broad field of machines performing tasks associated with intelligence | Planning, vision, speech, recommendations, and language |
| Machine learning | Systems learn patterns from examples instead of receiving every rule | Spam detection and product recommendations |
| Deep learning | Machine learning using multi-layer neural networks | Speech recognition and image understanding |
| Transformer | A neural-network architecture especially effective at relationships across sequences | Modern language and multimodal models |
| Foundation model | A large pretrained model adaptable to many downstream tasks | A base model used for writing, coding, search, or vision |
| Large language model | A foundation model focused on processing and generating language as token sequences | Drafting, extraction, classification, and conversation |
| Generative AI | Systems that create new text, images, audio, video, or code | Chat assistants, image generators, and coding assistants |
Module 4 · Apply it16 / 31
Apply the concept
Place six AI milestones on a timeline and explain what changed at each step.
- 1Place rule-based AI, machine learning, deep learning, transformers, foundation models, and assistant systems in order.
- 2For each milestone, state what the developer supplied and what the machine learned.
- 3Separate training from inference and pretraining from task-specific prompting.
- 4Explain why more capability can also create new cost, safety, and reliability risks.
Use this prompt
Teach the history from rule-based AI to modern LLM applications in seven milestones. For each milestone include: approximate period, central idea, familiar example, and one limitation. Avoid implying progress was a single straight line.
Success looks like
You can place LLMs inside the larger AI family and explain the difference between a model being trained and a product using it.
Watch for
- AI is the umbrella. Machine learning is one approach, deep learning is a machine-learning approach, and LLMs are one deep-learning model family.
- ChatGPT, Claude, and Gemini are products or model families, not synonyms for all AI.
Module 5 · Learn · 25 min17 / 31
LLMs, RAG, tools, and agents
An LLM is one component, not the entire AI product. Give it retrieved knowledge and you have a RAG pattern. Give it controlled functions and it can use tools. Let it choose and repeat actions toward a goal, with limits and checks, and you have an agentic system.
A real example
Asked about a refund, a plain LLM drafts a general answer. RAG first finds the company's policy. A tool-using workflow checks the order. An agent may gather evidence, select an allowed action, request approval, execute it, and verify the result.
Technical terms
- RAG
- Retrieval-Augmented Generation: find relevant trusted information and place it in the model's context before it answers.
- Agent
- A goal-directed system in which a model can choose among controlled actions, inspect results, and continue until a stopping rule.
Module 5 · Reference18 / 31
Choose the smallest useful pattern
| Concept | What it adds | Use it when | Main risk |
|---|---|---|---|
| Normal software | Deterministic rules | The outcome is known and repeatable | Rules become hard to maintain |
| LLM | Flexible language generation | The task needs interpretation or generation | Confident unsupported output |
| RAG | Retrieved trusted context | Answers depend on private or changing knowledge | Wrong retrieval produces wrong context |
| Tool workflow | Live data or controlled actions | Steps are known in advance | Permissions and integration failures |
| Agent | Chooses and repeats actions | The path varies and needs planning | Runaway cost, unsafe action, hard evaluation |
Complexity should be earned by a requirement, not added because a term is popular.
Module 5 · Apply it19 / 31
Apply the concept
Classify everyday AI products as a prompt, RAG flow, tool-using workflow, or agent.
- 1Begin with the smallest solution: normal software rules or one model request.
- 2Add RAG only when the answer requires private, current, or source-backed knowledge.
- 3Add a tool when the system must obtain live data, calculate, or change another service.
- 4Use an agent loop only when the path cannot be fixed in advance; add permissions, budgets, logs, approval, and stopping rules.
Use this prompt
Classify this AI product as: no AI, one LLM call, RAG, fixed tool workflow, or agent. Explain the minimum architecture, required data, risks, evaluation method, and why a more complex pattern is unnecessary.
Success looks like
You can choose the least complex pattern that meets the need and explain why an agent is not automatically better.
Watch for
- RAG does not retrain the model; it supplies relevant information at request time.
- Calling one API in a fixed sequence is tool use or automation, not necessarily an agent.
Module 6 · Learn · 30 min20 / 31
Choose the right AI model
There is no single best model. A useful choice satisfies the product's quality threshold while meeting limits for speed, cost, privacy, language, context length, modality, tool use, deployment, and licensing. Model catalogs change quickly, so test current candidates with your own cases.
A real example
A high-volume category label may need a fast inexpensive model. A difficult coding migration may justify a frontier reasoning model. Sensitive offline summarization may require a smaller open-weight model running inside the organization.
Technical terms
- Commercial API
- A provider hosts the model and charges for access, reducing infrastructure work.
- Open-weight model
- Downloadable model parameters available under a license; this does not always mean open-source training data or unrestricted use.
Module 6 · Reference21 / 31
Core language-model types
| Concept | Best for | Trade-off |
|---|---|---|
| Base or pretrained | Further training, research, or specialized adaptation | Usually not the right direct interface for an end-user product |
| General instruct | Writing, extraction, classification, support | Balanced rather than strongest at one specialty |
| Reasoning | Complex planning, analysis, math, difficult decisions | Usually slower and more expensive |
| Coding | Code understanding, generation, debugging, repository work | Must still be tested in the real codebase |
Module 6 · Reference22 / 31
Specialized model types
| Concept | Best for | Trade-off |
|---|---|---|
| Multimodal | Text plus images, documents, audio, or video | Input support and quality vary by modality |
| Realtime voice | Low-latency spoken interaction | Conversation quality, latency, and interruption handling matter |
| Embedding | Search, similarity, clustering, and RAG retrieval | Produces vectors, not a user-facing answer |
| Small or local | Privacy, offline use, edge devices, high volume | Lower capability or more deployment work |
Module 6 · Reference23 / 31
Representative current commercial families
| Concept | Provider | Good starting point | Consider when |
|---|---|---|---|
| GPT-5.6 family | OpenAI | General, reasoning, coding, tools, and multimodal products | You want a broad hosted platform and frontier models |
| Claude Opus 5 / Sonnet 5 | Anthropic | Long work, coding, professional tasks, and agents | You value strong instruction following and agentic workflows |
| Gemini 3.1 Pro / 3.6 Flash | Multimodal, long context, Google ecosystem, or high-throughput work | You need a Pro-versus-Flash quality, speed, and cost choice | |
| Grok 4.5 | xAI | Reasoning, coding, tools, and realtime information products | Its capabilities, ecosystem, and policies fit the product |
This is a dated snapshot. Availability and names can change, so verify the live provider catalog.
Module 6 · Reference24 / 31
Representative open-weight families
| Concept | Organization | Consider when | Verify first |
|---|---|---|---|
| Llama 4 | Meta | You need a broad ecosystem and deployment choices | License, hardware, and task quality |
| Qwen 3.6 / Qwen Coder | Alibaba Qwen | You need multilingual, coding, or varied model sizes | License and language-specific evaluation |
| DeepSeek V4 family | DeepSeek | You want cost-conscious reasoning or coding options | Current release status, hosting, data policy, and license |
| Mistral Small 4 and specialists | Mistral AI | You value efficient models, deployment control, or European hosting | License and whether a specialist or general model fits |
| Gemma / Phi families | Google / Microsoft | You need smaller local, research, or edge-friendly options | Device performance and task capability |
Open-weight is not automatically open source, free, private, or permitted for every commercial use.
Module 6 · Apply it25 / 31
Apply the concept
Select a model category and provider for four products, then defend each choice using measurable constraints.
- 1Define the task and create ten representative evaluation cases.
- 2Set required modalities, privacy boundary, latency target, and maximum cost.
- 3Shortlist one frontier model, one balanced model, and one low-cost or self-hosted option.
- 4Run the same cases, compare quality and operational fit, then pin the selected model version and monitor changes.
Use this prompt
Help me select an AI model for this product. First ask for task, modalities, languages, quality threshold, context size, latency, volume, budget, privacy, region, tool use, and hosting constraints. Then propose three candidates and a small evaluation plan. Do not choose from benchmark rank alone.
Success looks like
Your selection can be defended with test results and product constraints, not brand preference.
Watch for
- Provider names and model versions age quickly; verify the live catalog, pricing, availability, and deprecation policy before implementation.
- Open weights give control, not free operations. Include hardware, hosting, security, updates, and licensing in the cost.
Module 6 · Live references26 / 31
Model catalogs change
Verify the live provider source
This deck is a dated snapshot. Before choosing or deploying a model, confirm its current name, status, pricing, availability, context, modalities, and deprecation policy.
OpenAI model catalog
Current GPT and specialist models
Anthropic model overview
Current Claude tiers and selection guidance
Google Gemini models
Current Pro, Flash, multimodal, and specialist models
xAI models
Current Grok catalog and capabilities
Meta Llama
Official Llama family information
Qwen model catalog
Official Qwen organization and open-weight releases
DeepSeek models
Current hosted models and pricing
Mistral model overview
Commercial and open-weight model families
Model-family snapshot last reviewed: 31 July 2026.
Guided project27 / 31
Build brief
AI Lead Assistant · Stage 1
Translate the manual lead process into a software-and-AI blueprint, customize the reference business, and confirm the starter application works.
User story
As a visitor, I immediately understand what the assistant will help me decide.
Guided build · Part 128 / 31
Build it step by step
Map the software
Assign each manual step to the frontend, backend, API, database, AI, automation, or human.
Choose the business
Pick one service business and describe the customer decision the assistant will support.
Guided build · Part 229 / 31
Build it step by step
Write the promise
Complete: Answer a few questions and receive a clear recommendation for your next step.
Customize and run
Update the starter's business name and promise, then confirm it opens locally.
Save checkpoint one
Record the working version before adding the interactive journey tomorrow.
Ship checkpoint30 / 31
Do not ship until
- The manual lead process and its software equivalent can be explained without technical vocabulary.
- Frontend, backend, API, database, authentication, and cloud are correctly placed on the product map.
- LLM, RAG, tool use, and agent are distinguished with concrete examples.
- The chosen model category is supported by product requirements rather than popularity.
- The business theme, customer, and product promise are specific.
- The starter app runs locally and reflects the chosen business.
- No secret is stored in source code or GitHub.
Wrap and prepare31 / 31
Ship it. Show it. Prepare the next move.
Before the next day
- Prepare six realistic questions a business would ask a new lead.
- Write three example recommendations the assistant could return.
- Collect a logo or two visual references for the chosen business theme.