Hands-On Agentic AI Engineering — 31 Production-Grade Projects Built From Scratch

Author: Mahesh Guntumadugu · Live site: curiousengineers.in · Topics: AI Agents · Agentic AI · LLM Engineering · RAG · LLMOps · AI Observability · Multi-Agent Systems

A portfolio of 31 hands-on, production-grade AI engineering projects — real agents, real RAG pipelines, real LLM infrastructure — each built from scratch in Node.js/Python with minimal frameworks, tested, and hardened. Not slideware. Not notebook demos. Runnable systems that show how production agentic AI is actually engineered, from the ReAct loop up to observability, cost attribution, guardrails, and self-improvement.

---

For recruiters & hiring managers

What this repository demonstrates — end-to-end ownership of the modern Generative AI / LLM / AI-agent stack, at the depth expected of a Staff / Principal / Lead AI Engineer or AI Architect:

• AI agents & orchestration — ReAct loops, tool use, multi-agent coordination, durable long-running execution, self-improvement
• RAG & retrieval — hybrid search (BM25 + embeddings), reranking, chunking, eval-driven quality gates
• LLMOps & infrastructure — model routing, LLM gateways, cost/latency optimization, semantic caching, cost attribution
• AI safety & governance — prompt-injection defense, PII redaction, guardrails, compliance audit trails, EU AI Act model cards
• Observability & evaluation — decision-quality tracing, LLM-as-judge eval harnesses, drift detection, quality scoring
• Protocols & tooling — Model Context Protocol (MCP) server/client, AI coding agents, developer CLIs, voice agents
• Engineering discipline — automated tests, security hardening (rate limiting, CSP, graceful shutdown), CI/CD quality gates for AI

Core technologies: Node.js · Python · JavaScript · SQLite · Vector search · Server-Sent Events · React 19 · Model Context Protocol · Ollama / NVIDIA / Google Gemini LLM APIs.

Contact: 🌐 curiousengineers.in · 💼 LinkedIn · ✍️ Medium · 💻 GitHub · ✉️ [email protected]

---

The 31 projects

Every project is a self-contained, runnable system with its own README. Grouped by domain:

🤖 Agents & orchestration

#	Project	What it builds
01	PR Review Agent	A ReAct-pattern agent that reviews GitHub PRs and produces structured findings
03	Research Agent + Observable Harness	Instrumented agent loop with termination gates, convergence detection, JSONL traces
04	Multi-Agent Content Pipeline	Four specialized agents orchestrated by a supervisor to produce a blog post
16	AI Coding Agent	A mini Claude Code / Devin — reads a codebase, plans a fix, applies it, self-corrects on test failure
18	Agentic Workflow Engine	DAG engine: LLM nodes, tool nodes, approval gates, conditional branches, parallel fan-out
21	Multi-Agent Coordinator	Dynamic delegation, capability-based routing, escalation chains, real message bus
23	Long-Running Agent	Durable execution: checkpoint/resume, crash recovery, budget enforcement
25	Zero-Trust Agent Executor	IAM-style policies + sandboxed execution + human-in-the-loop for agent actions

📚 RAG, memory & context

#	Project	What it builds
02	Personal CRM Agent with Cross-Session Memory	Persistent knowledge graph across sessions — procedural, semantic, episodic memory
05	Codebase Q&A with Hybrid RAG	BM25 + embeddings, RRF fusion, reranking over a real codebase
09	Same Problem Three Ways	Prompting vs RAG vs Fine-tuning, evaluated head-to-head
22	Context Window Optimizer	Budgets heterogeneous sources (prompts, RAG, memory, tools) into a context window

⚙️ LLMOps, cost & infrastructure

#	Project	What it builds
06	Model Router with Cost Dashboard	Classifies query complexity, routes to cheapest capable model, escalates on failure
11	Optimize an Agent from $2 to $0.15	Four compounding optimizations that cut per-conversation cost ~90%
24	LLM Gateway — Routing & Governance	Routing, rate limiting, PII redaction, cost tracking between app and providers
27	Cost Attribution Engine	Cost-per-outcome across agents, waste detection, budgets, ROI — the CFO dashboard for AI

🛡️ Safety, evaluation & governance

#	Project	What it builds
07	Prompt Injection Test Suite + Defense	59 training + 29 held-out attacks across 5 categories, plus a defense layer
08	RAG Eval Harness	LLM-as-judge scoring across three dimensions with golden datasets
13	Bias Audit + Model Card Generator	Bias detection for AI resume screening, EU AI Act transparency docs
26	Agent CI/CD Pipeline	Eval suites on every PR, baseline comparison, quality gates, auto-promotion
28	Compliance-Ready Audit Harness	Tamper-evident event logging for regulated AI systems

📊 Observability, UX & protocols

#	Project	What it builds
10	SQL Analytics Agent with Permission Tiers	Text-to-SQL against a real DB with tiered permissions and error recovery
12	Production Chat UI with Trust Signals	Seven UX patterns — streaming, confidence, citations — that build user trust
14	FDE Customer Onboarding Toolkit	Forward-Deployed-Engineer toolkit for onboarding enterprise AI customers
15	MCP Server + Client	Hands-on Model Context Protocol — the tool-connectivity protocol behind Claude
17	Real-Time Voice Agent	Browser voice agent with interruption handling, turn-taking, live transcript
19	Agent Observability Dashboard	Traces, cost tracking, quality scoring, drift detection, live dashboard
20	aidev — AI Developer CLI	Zero-dependency CLI: commit messages, code review, file explanation via LLM
31	Agent Observability from Scratch	⭐ Flagship — decision-quality tracing (Tool ROI, coherence, confidence), tested & hardened

---

⭐ Featured: Project 31 — Decision-Quality Agent Observability

Most agent tracing tells you the run was slow. It can't tell you the run was dumb.

The flagship project reframes agent observability: instead of latency spans (the LangSmith/Langfuse model), it scores decision quality — Tool ROI (did each tool call earn its cost?), reasoning coherence, confidence signals, and strategy classification — persisted to SQLite and rendered in a run inspector. Pure Node.js, zero frameworks, 39 tests, production-hardened (per-IP rate limiting, CSP, graceful shutdown). → Read the deep dive

---

The interactive System Design Playbook

Alongside the buildable projects, this repo is a React 19 web app — an interactive playbook of production system-design decision trees (database selection, rate limiting, caching, message queues, scaling, event-driven architecture, resilience, observability, and more) plus a visual AI Engineering track with architecture diagrams for every domain above.

Run it locally

git clone https://github.com/gmaheshraju/Hands-on-AgenticAI.git
cd Hands-on-AgenticAI
npm install
npm run dev          # → http://localhost:5173

Each hands-on project runs independently — cd projects/<name>, read its README, npm install && npm test.

Stack: React 19 · Vite · React Router v7 · CSS custom properties (fluid clamp() typography, dark/light theme) · zero UI libraries · deployed on Cloudflare Pages.

---

Architecture gallery

---

FAQ

What is this repository? A hands-on portfolio of 31 production-grade agentic-AI engineering projects — AI agents, RAG, LLMOps, guardrails, evaluation, observability, and infrastructure — each built from scratch and runnable, plus an interactive system-design playbook web app.

Who is it for? Engineers learning how production AI agents are actually built; hiring managers evaluating depth in GenAI/LLM engineering; and anyone searching for concrete, framework-free reference implementations.

Is the code production-quality? The flagship (Project 31) ships with 39 tests and production hardening (rate limiting, CSP, input caps, graceful shutdown). Every project is a working system with its own README, not a snippet.

What makes it different from a tutorial? It builds the hard parts from scratch — the observability scoring, the RAG fusion, the LLM gateway, the guardrail defense — so you can read exactly how each mechanism works instead of importing a black box.

What is "decision-quality observability"? Scoring how good an agent's decisions were (did each tool call inform the answer? did reasoning stay coherent?) rather than only how fast they ran. Implemented from scratch in Project 31.

---

Keywords

agentic AI · AI agents · LLM engineering · generative AI · GenAI · large language models · RAG · retrieval augmented generation · LLMOps · MLOps · AI infrastructure · multi-agent systems · AI observability · agent observability · decision-quality tracing · LLM evaluation · LLM-as-judge · prompt injection defense · AI guardrails · AI safety · responsible AI · EU AI Act · Model Context Protocol · MCP · tool use · function calling · vector search · semantic caching · cost optimization · LLM gateway · model routing · self-improving agents · durable execution · zero-trust AI · AI coding agent · voice agent · Node.js · Python · React · system design · staff engineer · principal engineer · AI architect · forward deployed engineer · Mahesh Guntumadugu

License

MIT © Mahesh Guntumadugu