Project Titan is an active research deployment running on a single dedicated appliance on my physical desk. It explores what happens when autonomous agent loops, hardware energy constraints, and strict containment boundaries collide in real-world software engineering.
Rather than offloading compute to anonymous, remote megawatt server farms, Cindy Pawford lives on a localized edge appliance (ASUS Ascent GX10 / Apple Silicon ARM64) sitting directly in my workspace. In the dead of winter, dense local model inference and autonomous reasoning cycles literally heat my office. Every watt consumed dissipates as warmth into the room, converting software benchmarking into real thermodynamic utility.
Cindy does not monopolize the machine. Alongside sports analytics agents and primary operations orchestrators, the agents must learn to budget token allowances, serialize queue scheduling (enforcing max_parallel_requests: 1), and negotiate time slices on shared memory. They coordinate across a ~273 GB/s unified LPDDR5x memory bus without thrashing the host or starving peer agents.
Autonomy without boundaries is a liability. Here, we aggressively evaluate safety patterns: zero-trust host sandboxing where agent containers mount unprivileged directories, in-transit egress credential injection via Caddy proxies so agents never see raw API keys, pure human-auditable Open Knowledge Format (OKF) Markdown memory stores, and Closed Shadow DOM web shells that self-heal against hostile resets.
Cindy's primary external interface is Telegram. Community stakeholders can chat directly with her, propose weekly feature dispatches, or test agent conversational and creative limits. Autonomous sub-agents then synthesize approved ideas into live production code.
> Launch Telegram Dispatch (@CindyPawford_bot) →