I build AI-automation tools that ship — and I'm packaging them so you can ship too.

The pattern behind every monitoring/alerting bot is the same: scrape a source → rank or summarize with AI → alert you → run on a schedule. I run this stack across crypto signals, on-chain cluster detection, and OSS bounty hunting — and I'm turning it into tools you can pick up.

• ai-automation-skills (v0.6.0 — live) — 35 production-tested Claude Code skills for the scrape → AI → alert stack: Gemini cost control + prompt caching + structured output, Telegram + Slack alerts, cron scheduling, resumable pipelines, IMAP email digests, walk-forward validation + backtest comparison, OSS bounty + GitHub-label scouting, webhook intake, SQLite state, proxy rotation, vision extraction, hardened HTTP-retry sessions, CSV/Markdown reporting, .env config loading, rate-limit budgeting. One command: /plugin install ai-automation-skills@baronguyen001
• Trawlkit (shipping soon) — a paid starter kit that wires these skills into runnable Python bots: Playwright scraper + cost-controlled Gemini + Telegram/Discord alerts + scheduler + a tk CLI + 4 working bots. Follow @barobaonguyen for launch.

• gemini-agent-toolkit (v0.5.0 — live) — Gemini-native agent framework. Streaming, Google Search grounding helper, parallel/async tool calls, a golden-set eval harness, a gat CLI, selectable SQLite memory, cost tracking, prompt caching, structured output. 88% test coverage, CI on 3.11/3.12 × Ubuntu/Windows.
• gemini-cookbook (v0.5.0 — live) — 19 runnable google-genai recipes, no abstraction layer: prompt caching, structured output, batch API, cost tracking, thinking budget, function calling, search grounding, Files API, multimodal vision, embeddings, long-doc context caching, JSON-mode vs schema, safety settings. CI on 3.11/3.12 × Ubuntu/Windows.
• llm-prompt-linter (v0.2.0 — live) — ruff/eslint, but for your LLM prompt templates: catch injection-unsafe variable interpolation, hardcoded API keys, and token bloat before they ship. 12 rules across injection-safety / structure / cost / robustness; reads .txt/.md/.prompt files and .py prompt string literals; emits text / JSON / GitHub annotations / SARIF to gate CI (--strict, --min-score). Zero runtime deps, fully offline, never calls a model. CLI + library promptlinter, 98% test coverage, CI on 3.11/3.12 × Ubuntu/Windows.
• mcp-config-linter (v0.1.0 — live) — Your MCP config (claude_desktop_config.json, .vscode/mcp.json, Cursor, Windsurf) is a JSON file full of API keys and npx -y — lint it before you commit it. 19 rules across hardcoded secrets, unpinned supply-chain runners, shell-injection args, plaintext-HTTP remotes, and config hygiene → a 0-100 score, plus a redact command to safely share a config. Emits text / JSON / SARIF / GitHub annotations to gate CI (--min-score). Keyless, fully offline, zero runtime deps, never connects to a server. Python 3.11+ stdlib-only, 92% test coverage, CI on 3.11/3.12 × Ubuntu/Windows.
• skill-library-audit (v0.1.0 - live) - Every agent skill you install is taxed on every turn: its frontmatter name + description sit in context whether the skill fires or not. skillaudit budget ~/.claude/skills prices that per-turn bill and ranks the worst offenders (a 30-skill library measured 5,603 always-loaded tokens), skillaudit collide finds skills whose descriptions overlap so much the router picks the wrong one, and skillaudit audit runs 22 library-level rules (duplicate names, cross-scope shadowing, dead refs, leaked machine paths) into a 0-100 score with text / JSON / Markdown / SARIF output to gate CI. Existing SKILL.md linters check one file; this one asks what your whole library costs. Keyless, offline, deterministic, zero runtime deps. Python 3.11+ stdlib-only, 96% test coverage, CI on 3.11/3.12 x Ubuntu/Windows.
• news-vault (v0.1.0 — live) — Turn a daily news database into a password-protected static archive you can host on GitHub Pages for free: one HTML folder per day, AES-256-GCM payloads decrypted in the browser with WebCrypto, and a shell that contains no article text at all — cloning the public repo or reading the page source reveals only dates and counts. Diacritic-insensitive Vietnamese search across the whole archive, operators (source:reuters topic:AI impact:cao score:>70), deterministic same-story clustering, auto-generated entity and weekly pages, trending terms and coverage blind spots, build-time SVG charts, a ⌘K palette, saved articles, watchlist, offline PWA. Optional Gemini daily brief and one editorial illustration per category under a locked house style — both cached by content hash and both fail soft, so a dead API key costs you a picture, not the site. Incremental rebuilds keep the daily commit small. CLI + library, CI on 3.11/3.12/3.13.
• youtube-summarizer (v0.1.0 — live) — Paste a YouTube URL → a clean, format-aware summary in your terminal. Provider-agnostic (Gemini / Claude / OpenAI, your key), json3 caption dedup (~50% fewer tokens), SQLite retry intelligence, and stdout/Markdown/HTML/Telegram delivery. No browser, no login. CI on 3.11/3.12 × Ubuntu/Windows.
• playwright-cloudflare (v0.5.0 — live) — Playwright stealth defaults for authorized browser testing: explicit sync helpers, pw-stealth check (now with a CreepJS detection report), chrome/edge/brave/firefox fingerprint presets, and no extra stealth dependency. CI on 3.11/3.12 × Ubuntu/Windows.
• ai-news-aggregator (v0.5.0 — live) — Auto-summarize 35+ AI newsletters into one weekly digest. Gmail IMAP + RSS → Gemini classify + summarize → Telegram/Discord, with dedup, per-topic routing, a static HTML dashboard, and a built-in scheduler. ~$0.02/week on Gemini Pro.
• ideas-vault-kit (v0.4.0 — live) — A 4-pillar /40 scoring workflow to kill 9 of every 10 side-project ideas before you write a line of code. Opt-in AI score suggestions (BYO Gemini key), CSV/JSON + Obsidian-vault export, opt-in Notion sync, and an HTML GO/KILL leaderboard. CI on 3.11/3.12 × Ubuntu/Windows.
• walk-forward-validator (v0.5.0 — live) — Train/validation splitter for time-series strategy tuning: catch overfit before it reaches production. Now with Combinatorial Purged CV (CPCV) + a self-contained HTML report. Leak-tested, CI on 3.11/3.12 × Ubuntu/Windows.
• asset-allocator (v0.7.0 — live) — A keyless personal-finance toolkit: a 2-minute risk quiz maps to a target allocation across 7 buckets, then track daily P&L, drift, rebalancing, buy-only DCA contributions, custom targets, income/expense cash flow, milestone-year net-worth goals, and compound-growth projections. v0.6 adds a local web app (allocate serve) with editable detail tables, emoji icons, interactive Chart.js charts, and a net-worth-vs-goals trajectory — all offline. Static dashboard is localized EN/VI. Keyless prices (Stooq + CoinGecko). CLI + Flask web app, 87% test coverage, CI on 3.11/3.12 × Ubuntu/Windows. Not financial advice.
• ziwei-cli (v0.2.0 — live) — Compute a full Zi Wei Dou Shu (紫微斗数 / Tử Vi Đẩu Số) natal chart as structured JSON, not a screenshot: 12 palaces, soul/body stars, five-elements class, decade periods. TypeScript library + CLI, deterministic offline core (wraps iztro), with an optional OpenAI-compatible AI reading via an injectable client. 34 tests, 92% coverage, MIT. My first TypeScript package.
• salesops (v0.1.0 — live) — The boring-but-essential sales-ops chores in one keyless CLI: clean & dedupe lead lists, validate emails (typo domains, role/disposable inboxes, missing fields), apply opt-out suppression, ICP-rubric scoring to A/B/C tiers, segment, compute per-lead timezone send-windows, mail-merge drafts, and a weighted pipeline forecast with a self-contained HTML dashboard for the CEO. Works only on your own CSVs — never scrapes, enriches, or phones home. Zero runtime deps, fully offline, deterministic. Python 3.11+ stdlib-only, 96% test coverage, CI on 3.11/3.12 × Ubuntu/Windows.
• geo-auditor (v0.1.0 — live) — Will ChatGPT, Perplexity, Gemini, and Google AI Overviews cite your page? geo-audit check <file|url> scores it against 14 GEO/AEO heuristics across structure / machine-readability / authority / freshness, returns a 0-100 grade with a prioritized fix list, and generates a starter llms.txt. Keyless, offline, and deterministic — drop it in CI with --min-score. Grounded in the GEO paper (KDD 2024). CLI + library, 98% test coverage, CI on 3.11/3.12 × Ubuntu/Windows.

• confluence-scanner (v0.5.0 — live) — Walk-forward-first confluence scorer: TA + on-chain + funding (Binance + Bybit + OKX), with bull/bear/chop regime-split backtests and a static dashboard. The framework that rejects my own bad strategies. Package confscan, CI on 3.11/3.12 × Ubuntu/Windows.
• wallet-cluster-detector (v0.4.0 — live) — Detect wallet clusters buying the same token within minutes. Helius + DexScreener pipeline, score-weighted detector, paper-trade simulator, a Base/EVM read adapter, and a signature-verified webhook receiver. Package clusterdetect.
• dex-chart-dashboard (v0.2.0 — live) — Self-hosted DEX token chart dashboard: multi-timeframe candlesticks (1m–1d) from GeckoTerminal with 30m/2h resampled locally, live price/liquidity from DexScreener, and a throttle + 429 backoff + 15-min cache to survive the free tier. v0.2 adds read-only EMA/RSI overlays, a multi-token compare view, and PNG chart-snapshot export. No API keys. Flask + Chart.js, import/CLI dexscope. CI on 3.11/3.12 × Ubuntu/Windows.
• pump-research-framework (v0.3.0 — live) — A 4-layer (narrative / social / on-chain / catalyst) confluence framework + an offline deterministic scorer, case studies, and a sizing playbook. Now with a keyless narrative finder, a per-strategy backtest, and an HTML scorecard. Package pumpscore, stdlib-only.
• helius-rate-limiter (v0.6.0 — live) — Survive the Helius free tier: RPS limiter + monthly-quota budget + circuit breaker + multi-key rotation + token-bucket + a Prometheus exporter, plus an adaptive AIMD limiter and a restart-surviving persistent monthly quota. Sync + async, stdlib-only.
• pinescript-strategy-template (v0.4.0 — live) — Parameterized Pine v6 strategy templates (trend, mean-reversion, short-side) + a Python next-open-fill backtest engine with walk-forward replay/optimize, Monte-Carlo robustness bands (PNG export), holding-period sensitivity, and a TradingView "List of Trades" CSV parser. Package pinescript-walkforward.

Available for AI-agent / LLM integration, data pipelines, and scraping/automation work. Reach me: @barobaonguyen on X

Build-in-public threads on cost optimization, walk-forward validation, and shipping small products. See pinned tweet 👇