Ops · Internal Playbook
One goal: maximise your output per finite hour — cut the lag between you asking and it being done — so the team's exponential side compounds toward financial freedom.
transcribe.py (Whisper) and speak.py (Kokoro) reload the model on EVERY call — a multi-second tax on every voice message IN and every voice reply OUT. Highest-frequency, easiest to kill.
Idle heartbeats reload the growing conversation (one session: 119M cache-read tokens). Costs money AND slows every turn. "The loop, not the model, is the expensive part."
Railway ships every agent's files (forces strictly serial deploys); verify→render→re-render runs serial. A throughput wall on the whole swarm.
Near-$0, mostly self-hosted, hits all three root causes. A day or two combined.
Turn transcribe.py into an always-on faster-whisper server (model stays resident). Voice-in lag: seconds → ~0.5–0.8s. The single biggest hidden delay, on a tool we already own.
Same fix for Kokoro — keep the pipeline resident instead of reloading per reply. Cuts the lag on every voice reply out. Keep Kokoro (it's the right owned choice; Piper was archived Oct 2025).
Cache CLAUDE.md + MEMORY.md (the big repeated context). Up to 90% cost and 85% latency off long prompts — on every single message. Near-zero effort.
Haiku for simple replies / routing / extraction, Sonnet as workhorse, Opus only for hard reasoning + strategy. 60–80% cheaper and faster on the bulk of interactions.
Sub-second inference for latency-sensitive small tasks (incl. Groq Whisper as a hot STT fallback that offloads your Mac). $0 at your volume, no card.
Dead-centre on how you actually work.
See what the swarm is doing in real time, approve / reject / redirect, monitor multiple sessions — from your phone. Code + files stay on your Mac. A supercharged version of your Telegram-steering. ~15 min to switch on.
Serial → parallel delivery; run N agents at once and pick the best result. Free + native (we already use worktree isolation — lean in harder).
Audit found up to 3 overlapping inbound handlers running (the loop + tg_inbound_poller.py + tg_light_responder.sh) — risk of double-processing + wasted cycles. Collapse to one source of truth.
Split per-agent deploy scopes so deploys parallelise instead of queueing behind the "railway ships everyone's files" hazard. Removes a hard throughput wall.
Hot reads (Radar / leaderboard) behind KV, live per-entity state in Durable Objects, buffer on-chain ingestion through Queues. Cuts user-facing + pipeline lag AND advances the EyezOn owned-data engine. ~$5/mo floor.
Visual webhook / cross-app automation (Telegram ↔ Notion ↔ Gmail ↔ chain data) without bespoke glue each time. Self-hostable — owns the stack.
A small vector/temporal layer alongside our BM25 file-memory for facts that CHANGE (whale wallets, token status, pricing) — so recall never serves a stale contradiction. Low urgency; the file memory is already a genuine strength.
Gemini Live (~$0.005/min) skips the STT→LLM→TTS chain for live conversation — ideal for the AI voice-coach product. Claude Computer Use (~Q4 2026): "message a task from your phone, Claude drives the Mac" — turns GUI chores (Axiom, uploads, dashboards) into delegable tasks.
This week: #1–#5 together (a day or two, near-$0, all self-hosted/first-party) — they hit all three lag root-causes at once. Flip on #6 Remote Control today.
Then: #7–#9 for swarm throughput. The infra (#10–#11) rides the EyezOn-Data north-star as we build it.
Every item here reduces external dependency, not adds it — rented SaaS (Deepgram, ElevenLabs, realtime APIs) stays as hot fallbacks/bridges, never the foundation. Say which you want and I'll start — the ones touching live voice/inbound infra I'll only action on your go.
Synthesised from a full read-only audit of our stack + a broad 2025-26 tools sweep (STT/TTS, fast inference, automation, orchestration, mobile ops, memory, edge infra). Sources on file. All costs approx AUD.