My AI works
while I sleep.

Siti runs the day-to-day of the TODAK group over WhatsApp: staff ask her questions, she recalls company memory, files receipts, runs background jobs, and when a task needs real work she dispatches it to "hands" — live coding-agent sessions on actual machines. She has a standing rule to escalate rather than surrender: "I can't access that" is not an acceptable answer. The owner's messages run on a full agentic engine; everyone else gets a fast lane.

• Three brains: chat lane, background-task lane, and dev-task lane with real machine access
• Dispatches work to Claude Code sessions on fleet machines and reports back with receipts
• Escalation doctrine — blocked tasks route to a stronger lane instead of failing politely
• Speaks Malay and English, knows the whole company through shared memory

NACA is the platform under everything: an agent registry with heartbeats, command and intent queues, scheduled actions, and monitors that watch the monitors. Agents plug in declaratively — no hardcoded rosters — and satellite machines join the fleet the same way. When something dies, the fleet notices before Neo does.

• 36+ registered agents across VPS boxes, an edge GPU workstation, a Raspberry Pi and Macs
• Plug-and-play agent architecture — the registry is the single source of truth
• Fleet command center with live status, plus watchdog monitors validated end-to-end
• Multi-machine job routing: the right task lands on the right hardware

Agents without memory are goldfish. neo-brain is the fix: a central semantic memory where every conversation, decision, deploy and meeting gets embedded and becomes recallable by any agent in the fleet — Siti quotes decisions made months ago, coding sessions inherit context from each other, and meeting minutes sync in minutes after the meeting ends. It also holds an encrypted credential vault so secrets never live in code or chat.

• 8,800+ embedded memories with hybrid semantic + keyword retrieval
• One SDK for every agent — JavaScript and Python, with an audit trail per agent
• Encrypted credential vault: secrets stored once, referenced by name everywhere
• Knowledge-graph layer and meeting-minutes auto-sync feeding recall

Every day, The Forge checks how many new messages Neo has written. When enough accumulate, it builds a training set, scrubs anything secret, fine-tunes a local model on an RTX 5070, and scores the result for "Neo-likeness" against a held-out benchmark. Promotion is human-gated — the machine proposes, Neo approves. Feeding it requires nothing: just using WhatsApp. Current baseline: 80.2% Neo-likeness from 31,788 conversation pairs.

• Fully automatic loop: dataset build → secret scrub → train → eval → report
• Neo-likeness scoring gate — no silent regressions, no ungated promotions
• The trained model is served locally and powers a live agent today
• Zero manual data work: the training set grows by living normally

The twin answers as Neo, from Neo's real (public) memories — personality, facts, and history retrieved semantically per question. It has drafted WhatsApp replies in shadow mode, scored by the real Neo for authenticity, and its next upgrade — part of this site's rebuild — connects the public chat to the actual fine-tuned voice model The Forge trains, with live memory instead of an archive.

• Public chat that only sees memories explicitly marked public — privacy by architecture
• WhatsApp shadow-drafting with human scoring loop feeding future training
• Being rewired to the real fine-tuned twin model with cloud fallback