Portfolio/Projects/Personal Project
Case Study

O.N.T.E – Personal AI Ecosystem

Personal Project

Solo Developer

Overview

Built O.N.T.E (Operational Neural Thinking Engine) as a personal AI ecosystem running on WSL Ubuntu 24.04, integrating WhatsApp, Telegram, and Google Calendar into a single AI-powered pipeline — fully automated with human control as the override layer.

Designed a multi-layered message processing architecture: deduplication at ingestion, debounce buffering per sender, a global FIFO queue with 5–15 second inter-sender jitter, and per-JID rate limiting — all coordinated so AI responses feel human-paced and natural.

Implemented a human-in-the-loop approval flow via Telegram: when the AI detects scheduling intent embedded in WhatsApp messages, it routes parsed event details to the owner as an inline keyboard prompt (Confirm / Reject), then auto-inserts into Google Calendar on approval.

Built a full Telegram command center with a layered security model — guardCommand, guardMessage, and guardCallback functions at every handler boundary. Unauthorized access attempts are silently dropped or met with popups, while the owner receives automatic notifications with chat ID, username, and timestamp.

What Was Built

  • Implemented Baileys-based WhatsApp bridge with auto-reconnect, session persistence, and markOnlineOnConnect: false
  • Built debounce + global queue pipeline — sequential processing with realistic 5–15s inter-sender jitter
  • Designed AI-based schedule extractor using OpenRouter to parse natural-language messages into structured event JSON
  • Integrated Google Calendar API v3 with OAuth2 for automatic event creation post-approval
  • Implemented Redis context memory per contact — 10-message rolling window, 24h TTL, separate 7-day owner context for Telegram
  • Built whitelist v2 with name aliases, live disk reads (no restart needed), and natural-language management via AI
  • Designed dual system prompt architecture — limited WA persona vs. full-capability Telegram AI with no topic restrictions
  • Security hardened Telegram command center: per-handler guards, silent drops, and owner notifications with metadata

What's Next

In Development

Phase 1 — Agentic Action Layer

Transitioning O.N.T.E from a conversational system to an action-capable agent — able to execute real tasks directly from natural language, not just respond to them.

  • Structured intent-to-action pipeline replacing pattern-based detection
  • Real-time context awareness before responding to time-sensitive queries
  • Mid-conversation capture and routing to persistent storage
  • Unified automation gateway bridging all downstream integrations

Phase 2 — Workflow Automation Integration

Extending O.N.T.E's reach into real-world task execution — composing communications, generating documents, and orchestrating multi-step workflows from a single message.

  • Natural language to outbound communication — compose and send on behalf of the owner
  • Cloud document and structured note generation on demand
  • File handling with instant shareable link delivery back to chat
  • Automated daily intelligence briefings delivered proactively each morning
  • Event-driven follow-up reminders tied to calendar entries

Phase 3 — Persistent Memory Layer

Moving beyond session-scoped context toward a long-term memory system — enabling O.N.T.E to retain preferences, relationship context, and behavioral patterns across all interactions.

  • Embedded persistent store with zero operational overhead
  • Owner preference retention that compounds over time
  • Contact and relationship graph with contextual metadata
  • Behavioral pattern recognition across scheduling and communication habits
  • In-conversation memory read and write accessible to the AI at inference time

Phase 4 — Proactive Intelligence

Shifting O.N.T.E from a reactive assistant to one that anticipates needs — surfacing the right information at the right moment without being prompted.

  • Automated morning briefings with a full day's context delivered at a consistent time
  • Pre-event reminders with relevant details surfaced ahead of schedule
  • Weekly schedule previews initiated without user request
  • Passive monitoring for unattended communications with owner alerts

Phase 5 — Multi-modal Input

Expanding the input surface beyond text — enabling O.N.T.E to understand voice, images, and documents as naturally as written messages.

  • Voice input transcription and processing as a first-class interaction mode
  • Visual content comprehension for images and rich media
  • Structured data extraction from documents and receipts, routed to persistent storage

Phase 6 — Development Environment Bridge

Extending O.N.T.E into the software development workflow — enabling code-aware assistance and repository interaction through the same conversational interface.

  • Agent communication protocol research and architecture design
  • Repository analysis and code review accessible through natural conversation
  • Development task delegation with structured result summaries returned to chat
Nextarrow_forwardAI Customer Service SystemPT Textilindo