agentic-ai-sys

KO | EN

gitlite — search

#swift #go #anthropic #kubernetes #rag #cloud-native #redis #github #agents #proxy

(★ 15)

Agentic AI System Design with Case Studies

File Explorer

ZIP Download

- 01-see-the-mess.md
- 02-the-vocabulary.md
- 03-case-studies-insurance-recruitment-accounting.md
- 04-case-studies-real-estate-edtech-marketing.md
- 05-case-studies-hospital-support-legal-supply-chain.md
- 06-patterns.md
- 07-design-template-and-failure-modes.md
- 08-from-design-to-prototype.md
- 09-exercises-and-homework.md
.gitignore
README.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

**From Messy Workflows to Intelligent Agents**

*By Ayush Singh | 2026 Edition*

---

> *"Every business has a human doing 12 steps manually. Your job is to see those 12 steps and know which ones become ML, which become rules, which become LLM, and how they talk to each other."*

---

## About

A comprehensive, case-study-first guide to designing agentic machine learning systems. This book teaches you to look at real, messy business workflows and redesign them into clean, intelligent systems using four building blocks: Triggers, Deterministic ML, LLMs, and Tools.

No fluff. No theory without practice. Ten real industries. Seven named design patterns. One design template you can apply anywhere.

---

## Table of Contents

### Part 1 &mdash; Foundations

| # | Chapter | Description |
|---|---------|-------------|
| 01 | [See the Mess](chapters/01-see-the-mess.md) | The universal messy workflow pattern. How to ask the right 3 questions. Why humans can't see their own workflows. |
| 02 | [The Vocabulary](chapters/02-the-vocabulary.md) | The 4 component types (Trigger, Deterministic ML, LLM, Tool). The Golden Rule. Agent complexity levels 0-3. Orchestration and memory. |

### Part 2 &mdash; Case Studies

| # | Chapter | Industries | Key Lessons |
|---|---------|-----------|-------------|
| 03 | [Insurance, Recruitment, Accounting](chapters/03-case-studies-insurance-recruitment-accounting.md) | Insurance, HR, Finance | The Golden Rule in action. Embeddings for matching. Multi-agent parallel processing. |
| 04 | [Real Estate, EdTech, Marketing](chapters/04-case-studies-real-estate-edtech-marketing.md) | Real Estate, Education, Marketing | Memory + personalisation. Lead scoring. Creative generation with human review gates. |
| 05 | [Hospital, Support, Legal, Supply Chain](chapters/05-case-studies-hospital-support-legal-supply-chain.md) | Healthcare, E-commerce, Legal, Logistics | Triage classification. Sentiment-aware escalation. Citation requirements. Event-driven architecture. |

### Part 3 &mdash; Patterns and Practice

| # | Chapter | Description |
|---|---------|-------------|
| 06 | [Patterns You Should Now See](chapters/06-patterns.md) | 7 named design patterns: Triage, Enrichment, Match-and-Explain, Review Loop, Memory, Parallel Processing, Escalation. |
| 07 | [Design Template + Failure Modes](chapters/07-design-template-and-failure-modes.md) | The 5-step universal design template. 7 failure modes and how to design against them. The human-in-the-loop spectrum. |
| 08 | [From Design to Prototype](chapters/08-from-design-to-prototype.md) | 30-day build plans. Pseudocode for each component type. Metrics that prove it works. Build vs buy. |
| 09 | [Exercises + Homework](chapters/09-exercises-and-homework.md) | 3 live exercise scenarios. Homework submission template. Evaluation rubric. |

---

## The 10 Case Studies at a Glance

| # | Industry | Messy Workflow | Key Pattern | Complexity |
|---|----------|---------------|-------------|------------|
| 1 | Insurance | Claims processing (45 min/claim, 3 systems) | Triage + Escalation | Level 2 |
| 2 | Recruitment | Candidate shortlisting (6-10 hrs, 3-4 rounds) | Match-and-Explain + Memory | Level 2 |
| 3 | Accounting | Monthly expense reconciliation (3-4 hrs/client) | Parallel Processing | Level 3 |
| 4 | Real Estate | Property recommendations (3-4 rounds, 50% ghosting) | Memory + Personalisation | Level 2 |
| 5 | EdTech | Sales lead follow-up (70% time on cold leads) | Triage + Escalation | Level 2 |
| 6 | Marketing | Campaign creative workflow (2-3 weeks to approval) | Review Loop | Level 2-3 |
| 7 | Hospital | Appointment booking (wrong dept 20%, no triage) | Triage + Compliance Boundary | Level 2 |
| 8 | E-commerce | Customer support (2K-3K tickets/day) | Triage + Escalation + RAG | Level 2 |
| 9 | Legal | Contract due diligence (200-500 docs, 2-4 weeks) | Parallel Processing + Citation | Level 2-3 |
| 10 | Supply Chain | Inventory and demand planning (5 spreadsheets, weekly) | Event-Driven + Predictive ML | Level 3 |

---

## The 7 Design Patterns

| Pattern | One-line | Used In |
|---------|----------|---------|
| **Triage** | Classify first, route second | Insurance, Hospital, E-commerce, EdTech |
| **Enrichment** | Fetch from multiple sources, merge, then decide | Insurance, Accounting, Legal, Supply Chain |
| **Match-and-Explain** | ML scores, LLM explains why | Recruitment, Real Estate, EdTech |
| **Review Loop** | Generate, human reviews, feedback, regenerate | Marketing, Legal, Recruitment |
| **Memory** | Store feedback, change future behavior | Real Estate, Recruitment, EdTech, Supply Chain |
| **Parallel Processing** | Split across agents, reconcile at end | Accounting, Legal, Supply Chain |
| **Escalation** | Auto-handle easy, route hard to humans with context | Insurance, E-commerce, Hospital, Supply Chain |

---

## The Golden Rule

> **The LLM never makes the final call on anything that costs money, has legal consequences, or affects someone's health. The deterministic ML model or a rule-based system makes that call. The LLM only drafts the language after the decision is made.**

---

## Author

**Ayush Singh**

---

## License

agentic-ai-sys

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

agentic-ai-sys

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README