roy-vinay/applied-ai-agents

Working AI agents for real business problems, each with the guardrails and evals that make it
trustworthy, and the product decision behind it written down.
Every project here starts the same way: a real problem, the smallest version that works, evals on the
behavior that matters, then a write-up of the trade-offs. Clone any folder, run it, and pull it apart.
Run one now
The fastest start needs no API key and no dependencies:
Production patterns
The parts that keep an agent safe once real users and real money are involved.
| Project | What it shows | Stack |
| --- | --- | --- |
| Guarded support agent | Authorization in code, retries with backoff, step and token budgets, idempotent writes, PII redaction, prompt-injection screening, 12 trajectory evals, and a fault-injection benchmark, all in CI. Companion to Building AI Agents That Survive Production. | Python, no dependencies |
**What the guards prevent** under fault injection, 100 randomized trials per row, same agent with guards off vs on
(method and caveats):
| Outcome | How it arises | Guards off | Guards on |
| --- | --- | ---: | ---: |
| Duplicate payment after a lost response | Injected environment fault | 79/100 | 0/100 |
| Refund paid on another customer's order | Input: customer names any order | 47/100 | 0/100 |
| Refund over $50 paid without approval | Input: any eligible order | 25/100 | 0/100 |
| Payment made after an injection attempt (attack success) | Adversarial input | 68/100 | 18/100 |
| Card number shown back to the customer | Injected model failure, every trial | 100/100 | 0/100 |
| System prompt revealed | Injected model failure, every trial | 100/100 | 0/100 |
| Turn ran past 8 model calls | Injected model failure, every trial | 100/100 | 0/100 |
| Legitimate refund wrongly blocked (false positive) | Ordinary requests | 0/100 | 0/100 |
Agents
| Project | What it does | Live | Stack |
| --- | --- | --- | --- |
| MCP chat | Terminal chat app showing the Model Context Protocol end to end: an MCP client and server over stdio, with @document mentions and /commands. | | Python, MCP |
Tools
Tools for people who work alongside AI agents.
| Tool | What it does | Stack |
| --- | --- | --- |
| mdEditDiff | Free, open-source (MIT). An editor for people, a diff for agents. Agents ask what changed in a shared Markdown file since they last looked and get back only the changed parts, word by word, with who reported them, instead of re-reading whole plans. People get a native Mac app that reads and edits Markdown, with Read, Edit and Split modes and saves that never touch what they didn't change. Download the Mac app or install the CLI and MCP server. | TypeScript, MCP, Swift |
Failure Files
Case studies with experiments: a documented public AI incident, the failure mode behind it reproduced in a
fictional system, and mitigations measured one at a time. All files and method.