AI engineer. I build agentic systems and the infrastructure to trust them: permission-aware retrieval, agent observability, human-in-the-loop approval, and evals that fail CI when behaviour regresses. Building continuously.
Open source: Google Summer of Code, Linux Foundation (LFX), and Symmetry Autumn of Code. 3+ years shipping in production.
Agent infrastructure
- Tenantq: Multi-tenant hybrid search on Qdrant: payload tenant isolation, dense+sparse RRF, HNSW tuning, and benchmarks.
- VaultRAG: Permission-aware RAG: access control enforced inside the retrieval query, CI fails on any leak.
- MCP-audit: Security scanner for MCP servers: 18 rules, SARIF output,
npx mcp-audit <target>. - Skill-audit: Security scanner for agent skills: 31 rules, prompt-injection and exfil detection, SARIF output,
npx @royalpinto007/skill-audit <path>. - Answerproof: Verifiable receipts for RAG answers: Merkle inclusion proofs and Ed25519 signatures.
- Agentrace: Observability for Claude Code subagents: flags the agent results you should not trust.
- Ctxlens: Context-window profiler for agents: shows what is eating your tokens and cuts the waste.
- Ctxtrim: Trims AI-context bloat: finds the files ballooning your Claude Code, Cursor, and Codex token cost and writes the ignore files to cut it.
npx ctxtrim. - Evalgate: Prompt regression CI: the build fails when your prompt gets dumber.
- Voiceeval: Evaluation for voice agents: catches the failures a text eval scores as a perfect call.
- Injection-arena: A self-hostable prompt-injection challenge game: beat stacked AI defenses, climb the leaderboard.
- Tracecase: CI for AI agents: replay runs against prompt changes, catch regressions before they ship.
- Bridgekit: Scoped MCP server with per-client permission boundaries and an append-only audit log.
- Casebook MCP: Remote MCP server over AgentPostmortem, a public registry of real AI-agent failures.
- Casebook Chat: Streaming chat that investigates agent failures and answers with cited case IDs.
Agents in production
- Resolvd: End-to-end inbox operator: auto-resolves within policy, escalates the rest with a proposed action.
- Greenlite: Approve your agents from your phone: one-tap human-in-the-loop for escalations.
- Webhands: Computer-use agent for tools with no API: refuses any write unless you confirm it.
- RelayG: Support triage as a LangGraph state machine, paused mid-run by a human-in-the-loop interrupt.
- AgentPostmortem: Every AI agent failure, documented: a public case registry.
Systems & apps
- Kabootar: Offline mesh messenger: chat with no internet, no servers, no SIM. Messages hop phone-to-phone over Bluetooth and Wi-Fi (delay-tolerant network, epidemic routing, end-to-end acks). Flutter.
- Tiny Day: A cozy, offline-first daily planner for Android: gentle scheduling, focus, replanning, no accounts.
- Auralis: A real-time interactive aurora rendered on the GPU in the browser, WebGPU with a WebGL2 fallback, no server. Move your cursor to bend the light.
- Mycelia: A GPU slime-mold simulation in the browser: 262k agents grow living networks and flow toward your cursor. WebGL2 GPGPU, no server.
Products
- SignalizeAI: AI prospecting as a Chrome extension, featured on the Chrome Web Store.
- PennyRush: A private money hub: receipts and statements parsed in memory, only extracted fields stored.
Awesome lists
- distro-skills: 26 Agent Skills that teach an AI agent to distribute a dev/indie product (Show HN, Product Hunt, awesome-lists, dev.to, GitHub, Reddit, X, npm/PyPI). The builder's counterpart to marketing skills.
- awesome-agent-evals: Curated, auto-updating directory of tools, frameworks, and benchmarks for evaluating AI agents. Filterable site.
- awesome-llm-guardrails: Curated, auto-updating directory of open-source LLM guardrails: injection defense, PII, content safety, structured output, red-teaming.
Roots
- cpp: The code I wrote while learning C++ at the very start, kept detailed on purpose: datatypes through graphs and DP.