01 / RECRUITING AI
AI Smart Career Portal
A full recruitment platform where AI does the heavy lifting: it screens and shortlists CVs, scores candidates on hiring pillars, generates job descriptions, and runs automated voice interviews end-to-end.
AI features
- AI CV screening & shortlisting — keyword extraction + LLM evaluation + semantic similarity via OpenAI embeddings (text-embedding-3-large).
- AI voice interviews — Whisper transcription + GPT scoring on relevance, technical accuracy, clarity & completeness.
- 5-pillar candidate scoring — bias-free LLM ranking on commitment, communication, stability, learning & attitude.
- Auto job-description generation across 11 industry categories with role auto-detection.
- Combined score = 50% LLM + 30% semantic similarity + 20% keyword match.
Laravel 8React 18 + TS + ViteFastAPIOpenAI / GroqWhisperRedisi18n (5 langs)
02 / RAG CHATBOT
Conversa AI — Self-Training Support Bot
An enterprise support chatbot that trains itself from a knowledge base using a full RAG pipeline, answers customer questions with grounded responses, and can even handle apartment bookings through chat.
AI features
- End-to-end RAG pipeline — dataset ingestion → chunking (800 chars / 120 overlap) → OpenAI embeddings → vector indexing.
- Vector database — Qdrant (primary) or Pinecone, with candidate → stable index promotion.
- Incremental auto-training via background jobs; new docs ingested through webhooks.
- Budget-aware LLM routing — per-company rate limiting + graceful fallback when limits are hit.
- Conversational booking — apartment reservations completed inside the chat.
- Multi-company isolation with permission-scoped access.
Laravel 12Qdrant / PineconeOpenAI GPT-4o-miniLangChainNext.js 16Pusher (real-time)
03 / MULTI-AGENT
Master Agent Group — An Office of AI Agents
A multi-agent team that runs like a software company: a Business Analyst agent, backend & frontend developer agents, and a QA agent. I assign daily tasks, the BA logs them into Jira, assigns developers, QA verifies, and work is delivered for human approval.
AI features
- Specialized agent team — 3 backend + 2 frontend dev agents, plus BA & QA agents working as a pipeline.
- Jira-driven workflow — agents create tickets, transition status (To Do → In Progress → In Review), and assign developers.
- Full delivery loop — BA writes the spec, developers implement, QA checks, work is delivered.
- Safety gating — only processes tickets labeled
ai-agent; humans approve before Done.
- Working-hours aware — scheduled runs (Sun–Thu, office hours) via cron.
CrewAIClaude Agent SDKPythonJira REST APICron
04 / DEV AUTOMATION
Emergency Telegram Bug-Fix Bot
For emergencies: send the bot a screenshot of a bug with a caption, and it resolves the issue automatically — classifying the right repo, writing the fix, and opening a GitHub pull request for review.
AI features
- Image-based bug classification — Claude reads the screenshot + caption and routes to the correct repository.
- Two-stage LLM — classifier (pick repo) → fixer (implement the patch) via headless Claude.
- Auto pull request — never pushes to
main; opens auto/fix-* branch + PR for human review.
- Per-repo memory — optional CLAUDE.md cuts fix time by ~80% (1–3 min warm).
- Cost-efficient — roughly $0.10–0.50 per fix; runs as a systemd service.
Claude Agent SDKpython-telegram-botGitHub CLISystemd
05 / AI-DLC
AI Development Lifecycle (AI-DLC) System
A complete AI-driven development lifecycle integrated into a Laravel + React product. A phase-driven process — from discovery and specs to design and implementation — coordinated across multiple repositories with Jira and custom Claude Agent skills.
AI features
- Phase-driven workflow — discovery → spec → design → finalize → implementation, tracked per feature.
- Cross-repo coordination — Claude auto-routes work to the correct product repo (backend / client / media / mobile).
- Custom Claude Agent skills & slash commands:
/new-feature, /fix-bug, /refactor, /review, /security-review.
- Jira worklog with per-phase time tracking and a live progress dashboard.
- Brownfield mode — reverse-engineers a running system into a clean SRS.
Claude Agent SDKLaravel + ReactJira APIPython tooling
06 / WORKFLOW AI
n8n AI Automation — Lead Capture & Auto-Reply
No-code/low-code automation workflows powered by AI: Messenger lead collection where an LLM qualifies and responds to incoming leads automatically, plus a family of AI-response products built on n8n.
AI features
- Messenger lead collection — AI captures, qualifies, and stores leads automatically.
- AI auto-response — context-aware replies generated by an LLM in the flow.
- Composable automations — reusable n8n nodes that plug AI into any business process.
n8nLLM APIMessenger APIWebhooks
07 / DAILY INTEL
Daily News & Learning-Plan Telegram Bot
A personal AI assistant on Telegram that delivers a daily briefing — curated news plus a tailored learning plan — keeping me on track every single day. Part of a broader async "AI Engineering Life OS."
AI features
- Multi-agent system — briefing, market-scout, planner & reflection agents working together.
- Daily delivery — news digest + personalized learning plan pushed to Telegram.
- Cost guardrails — per-agent budgets, model downgrades, and hash-based caching to avoid duplicate LLM calls.
- Fully async pipeline — FastAPI + Redis/ARQ job queue.
FastAPIlitellmRedis + ARQpython-telegram-botMySQL
08 / AI DEVOPS
Chat-Driven AI DevOps
A CLI agent that turns natural-language commands into provisioned, hardened production servers. Say "deploy this" or "diagnose the server," and the agent detects your stack and runs audited Ansible playbooks.
AI features
- Natural-language ops — provision Laravel, React, or Next.js stacks by chatting.
- Stack auto-detection — reads composer.json / package.json / .git to pick the right playbook.
- Safety model — read-only diagnostics run freely; every mutation needs explicit y/n confirmation.
- Secrets handling — generated credentials stored at mode 0600, shown once.
Claude Agent SDKAnsibleTyper + RichPython 3.10+
09 / BROWSER AI
AI-Assisted Chrome Extension — Daily Chairman Report
A Chrome extension that extracts data from internal software, analyzes it, and delivers a daily update straight to the chairman — automating reporting that used to be manual.
Features
- Data extraction & analysis — pulls work-order / PO data from dashboards and summarizes it.
- Automated navigation — searches, opens reports, and renders PDFs via the Chrome debugger API.
- Queue + status tracking — CSV-driven processing with live status and auto-export.
- Daily delivery — packaged report sent automatically.
Chrome Extensions (MV3)JavaScriptService WorkerDebugger API
10 / TASK ENGINE
Every-Task — Daily Intake → Parallel Agents
A daily markdown file (one task per line) is auto-parsed into Jira tickets and processed in parallel by the right agents, with phase-dependency rules so multi-step features run in the correct order.
AI features
- Markdown → Jira → agent queue — write a line, get a ticket and an assigned agent.
- Phase-dependency DSL — features run in parallel; phases within a feature wait for prerequisites.
- Auto-routing — tasks dispatched to free agents of the correct role (BA / backend / frontend).
- Dry-run preview before any tickets are created.
Claude Agent SDKPythonJira APIMarkdown DSL
11 / VISION AI
AI Image Scanner — Auto Title & Description
A multimodal pipeline that scans any image and automatically generates an accurate title and a rich description — powered by vision-capable LLMs through OpenRouter, with a self-hosted Ollama fallback for private, zero-cost local inference.
AI features
- Image understanding — vision LLMs "read" the image and extract subject, context, and key details.
- Auto title + description — generates SEO-friendly titles and structured descriptions from a single image.
- OpenRouter routing — access to many vision models behind one API, with model fallback and cost control.
- Self-hosted Ollama — run open vision models locally for privacy and zero per-call cost.
- Hybrid strategy — cloud for quality, local for cheap/offline — switchable per request.
- Batch-ready — process many images in one run for catalogs, listings, or media libraries.
OpenRouterOllamaVision LLMsPythonMultimodal