Learn, build, and share AI engineering with the AI Engineer Help community.
This repository brings together independent examples, experiments, and learning resources. Each project lives in one primary category and owns its dependencies, documentation, and verification steps.
The directories follow the community’s learning categories; agents/ is the shortened name for Agents & Workflows. Each category README explains its scope and indexes its projects.
General and Site Feedback remain community discussion spaces rather than code directories. Product Breakdowns maps to showcase/, matching its community URL slug.
Building a Multi-Agent Travel Planner with Gemini and TypeSafe AI
Travel Lab compares Gemini-only coordination with Jev agent and tool selection, using the same Gemini workers and limits. The TypeScript website researches live places with Gemini Google Search grounding, shows decision traces and evidence, and saves each run. Synthetic catalogs remain for offline tests and evaluation. Local web-search smoke comparisons completed for both teams; controlled routing and whole-agent benchmark results are included. Public-deployment review remains outstanding.
Incident response agent investigates operational incidents from logs: a LangGraph agent plans, queries 14 public Loghub datasets, cites evidence, proposes fixes behind guardrails and human approval, and writes a root-cause report. A Next.js dashboard shows its live activity, the logs, a knowledge graph built from them, and an evaluation against expert labels (27 of 28 BGL incidents right, 28 of 28 on unseen cases). It runs locally with Docker; the README describes how it works and ways to extend it.
Dino Runner × Jev (live demo) is an original endless-runner game in the style of Chrome's offline dinosaur. Code runs the physics and timing, and Jev chooses one move per obstacle. Real-time and paused modes show how decision latency, not just correctness, decides the outcome, and a live telemetry panel plots every call's round trip against the time the game allowed. Tests and the benchmark dry run need no API key.
Biomedical hybrid search compares BM25, dense search, and hybrid retrieval over biomedical passages, with optional Gemini or TypeSafe Jev reranking, with evidence inspection, Gemini-generated answers, and a fixed 100-question evaluation set. Included benchmark results are explicitly labeled as a historical local-model baseline.
LLM-as-a-Judge: Measuring agreement, bias, and reliability (read the report) tests Gemini and Jev judges on AI-written synthetic mistakes and on 120 real answers from RAGTruth, whose errors were marked by people. Judges score 98–99% on the easy set but 53–68% on real answers, and a review of the disagreements shows the labels are often the ones at fault. A local dashboard shows every answer next to every verdict.
Guardrails for AI agents: what each layer actually stops (read the story) teaches guardrails through an illustrated story, The Gatekeepers of Lantern Mountain, then measures them on a refund-support agent on 115 labeled messages and 26 attack and benign scenarios. Gemini and Jev classifiers caught 39 of 40 attacks where regex caught 8, but with a hijacked model, detection still let 5 of 17 harmful actions through; authorization and business rules in code stopped all 17. Tests and the hijacked-model runs need no API key.
Turning policy into tests: Giskard for agent evals in finance and healthcare explains how legal, compliance and clinical teams can write the rules an agent must follow, and how the open-source Giskard library turns each rule into adversarial test conversations and an LLM judge. It walks through a five-step process on a banking assistant, then applies the same method to a patient-facing clinic assistant. A demo app runs real Giskard against a banking and a clinic agent on Gemini. You pick a risk, watch the attacker try to break the rule over several turns, and read the judge's verdict. In recorded runs, writing the policy into the prompt took investment-advice attacks from 0 of 10 passing to 10 of 10, and medication-change attacks from 1 of 10 to 9 of 10. Recorded runs replay without an API key.
Prompt caching: what breaks it, what it saves, and what we measured (try the playground) explains how Claude, Gemini and OpenAI reuse a prompt's prefix. It shows eight layout mistakes, such as a timestamp at the top, that switch caching off without an error. An interactive playground highlights those mistakes in your own prompt, estimates the cost per 1,000 requests, and reorders the prompt in one click. A live experiment on Gemini 3.8 Flash found that a misplaced timestamp cut the reusable prefix from 99.7% to 14.4%. Implicit caching hit 0 times in 97 requests, while explicit caching hit 12 of 12 and cut the cost per request by 84%.
Chunking by document type (try it) is a small web app: pick a document type (API docs, a FAQ, a research report, Python code, a CSV, a contract, a meeting transcript, a report with charts, a slide deck or a manual with diagrams) and it highlights how that document is chunked. Each type shows where we cut, what we keep together and the context we attach to every chunk, with a fixed-size split for comparison. It runs in the browser with no API keys.
Tools, tool calling and unsafe execution (try it) is a small web app about the one step where agent safety is won or lost: your code running a tool the model asked for. Pick one of five tools (read_file, fetch_url, run_command, calculate, lookup_order) and a call the model might make, and watch a naive implementation and a safe one run side by side — the naive one reading a file outside its folder, fetching an internal address, running an injected shell command, evaluating arbitrary code or leaking another customer's order, each shown next to its fix. It maps the five bugs to OWASP's Top 10 for Agentic Applications (2026) and to real CVEs (EscapeRoute, mcp-remote). All simulated in the browser, no API keys.
Anatomy of a harness (try it) draws an agent harness as an animated isometric factory on one full-screen page: the model sealed in the middle, the context window as a tray in front of it, and eleven numbered parts around it for context management (skills, memory, summarization, offloading), the execution environment (REPL, sandbox, file system, write and run code), delegation (planning, subagents) and steering (human in the loop). Watch crates of work move between stations as it fixes a failing test, then pull parts out and run it again: without subagents the run bills about four times the tokens, and without subagents, offloading and summarization it runs out of window. Every number comes from one scripted run; no API keys.
Pharma research agent is a production-shaped research workspace for drug and clinical-evidence questions. An orchestrator reads a skills layer and plans the research, parallel sub-agents search PubMed, ClinicalTrials.gov, FDA labels and the web, and a synthesis agent writes a report with every claim cited and checked. It's a Next.js 16 app with durable Vercel Workflow runs, Postgres, live progress, PDF and Word export and share links. It runs offline with no key; live research needs a Gemini key.
Wiki memory for agents (try it) explains the pattern Andrej Karpathy calls an "LLM Wiki". Instead of re-searching raw documents for every question, an agent keeps its own notebook of markdown pages and updates it as each new document arrives. The app starts from one question that plain retrieval answers with three conflicting passages. It then explains the idea and lets you step through five documents: for each one, what it says, what the agent did, and the page change that matters most. A check catches values that disagree, and the same question is answered again from one page. The pages are scripted; everything derived from them is computed. No API keys.
Knowledge graphs for AI engineers explains what a knowledge graph is, how an LLM extracts one from text, how entity resolution merges names like "Varnell Semi" and "Varnell Semiconductor", and the GraphRAG retrieval patterns. Gemini extracted a graph from 25 synthetic documents three times, rules, an LLM and a person merged the names, and Neo4j answered 15 questions: on the multi-hop question chunk search found 0 of 4 customers and Gemini-written Cypher 4 of 4, while on a prose question chunk search won. The app replays the recorded runs, real model mistakes included, with no keys; locally, with Neo4j and a Gemini key, it also answers your own questions.
LLM vs RAG vs AI agent vs agentic AI explains the four terms as layers, each the previous one plus one capability: retrieval, tools in a loop, then coordination of several agents. Animated figures show each level as a bench, short Python snippets show RAG and an agent loop, and a table helps you pick the lowest level that does the job. The snippets' tests use a mocked model and need no API key. Written by Nagaraju Poralla.
Clinic voice agent is a phone receptionist for a fictional clinic, built on LiveKit Agents and Gemini the way a voice platform ships one to a business. It looks the caller up in a CRM before answering, books and moves appointments with tools, answers from a knowledge base, and hands emergencies and requests for a person to code-level guards that transfer the call before the model is asked. After the call it writes a record from the tool calls and posts it by signed webhook. Talk to it in the browser over a realtime (Gemini Live) or cascade (STT → LLM → TTS) pipeline, break its CRM mid-call, and read each call's trace: latency by component, tool calls, retries and cost per minute. Twelve simulated callers, including a lost booking response that must not book twice, run offline with no key.
Context Compression Lab teaches four ways to cut context, step by step, on one handbook question whose answer depends on two facts. Click through lexical relevance, MMR, word deletion and prefix truncation with every number computed live, watch each one break (a reworded question, a stronger repetition penalty, a dropped “not”, a cut that never reaches the refund policy), then compare all four on one budget slider. Cards explain what model methods (BGE, LLMLingua-2, LLMLingua, LongLLMLingua, Laya, Compresr) change, and a lab takes any text. It runs in the browser with real cl100k_base token counts and no API keys. Live: context-compression-pink.vercel.app.
- Choose the category matching the main lesson of your contribution.
- Create
<category>/<project-name>/and document how to set up, run, and verify it. - Add a link to the category index. Cross-link from other categories when useful instead of copying code.
- Link related community discussions from the project README and share the code link in the discussion.
Use community tags for providers, frameworks, languages, and topics that span categories. Keep project tests and deployment files with their project.
Read CONTRIBUTING.md to add a project and AGENTS.md for repository conventions. CLAUDE.md imports the same instructions for Claude Code.
There is no shared installation or test command yet. No repository license has been selected.