Thirty Roles, One Model: Fortune 50 in a box

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

30 named AI roles. One 8B model. Zero cloud. A full reporting hierarchy, plus manuals, actor memory, local speech, and a rank-gated factory board, running on a single desktop GPU. This paper documents TeamChat V2: thirty seats, each bound to an actor and a reporting line, served by one locally hosted eight-billion-parameter instruction model inside one Windows process. No network calls for inference. No per-agent endpoints. 4-bit NF4 quantization packs the model into 5.4 GB allocated / 5.6 GB reserved on a 20 GB GPU at load. A three-stage router still decides who answers: explicit mention, one short greedy generation, or a compiled fallback that cannot operate hardware. Selected roles answer on a streamed single-shot path, a batched path, or a one-generate tool loop. Managers may still delegate to direct reports. Both recorded sessions compiled those last two paths and recorded zero calls on them. V2 adds the layers the chat-only system did not ship. Project knowledge and six short operating manuals. Per-actor work, memory, and knowledge files. Optional skill files. Local Windows speech from the actor catalog. Rank-gated tools on two channels: a short-lived executable for climate, alert, and spider release; Python serial for decorative lamps and factory polls. An Arduino factory that keeps running on USB power. A mapped remote key is an interrupt, not a preset. External web tools exist in source and are compiled off. The measurement layer writes eight append-only JSONL streams. An aggregator turns those logs into the report, provenance manifest, and LaTeX tables. Every number below is the converter output from a recorded session. Nothing was typed by hand. Key measured results, tool-run session (33 messages: 32 user, 1 factory interrupt): Model load: 11.7 s Allocated device memory at load: 5.4 GB (peak allocated at load 5.5 GB) Average routing cost, when a router generation ran: 1.7 s / 18.3 completion tokens Single-shot decode: 14.18 tokens/s Mean time-to-first-token (streamed simple answers): 1.46 s Full user turn (routing + every role answer): mean 13.7 s 18 tool calls, 161 hardware actions, 70 speech clips: all recorded successful Key measured results, monitor-run session (156 autochat messages, mention-routed): Model load: 11.7 s Single-shot decode: 19.82 tokens/s Mean time-to-first-token (streamed simple answers): 1.73 s Full autochat turn: mean 21.0 s 57 tool calls, 872 hardware actions, 336 speech clips: all recorded successful