Here is the whole server, one line at a time

Domain-agnostic trend-detection pipeline: ingest → extract → resolve → score → interactive graph predictor-suite → studio:predictor →

Cytoscape-based living map of a field — provenance on every edge explorer →

Daily curated ingestion + LLM extraction behind quality gates, batch-API economics (~$0.012/doc) ingest →

Multi-agent platform: custom crew running a real publishing operation with shared decision trace agents →

Director agent — coordinates the crew, human approval in the loop, per-run cost logged agents#director → see also: agent-platform.decision-trace

Reads the server's own health every morning at 06:20 and writes up what it finds — proposes, never applies agents#sysadmin → see also: ops.truth-map-practice

Wanders the box unprompted at 05:45 looking for stories nobody assigned — cheap model to walk, premium model to make the leap agents/newsroom#scout → see also: agent-platform.scout.leads-ledger

Drafts field notes from a claimed lead, in a voice loaded from bottled exemplars rather than described in adjectives agents/newsroom#writer →

Triage desk for inbound wire items, with a chief's shadow pass scoring the desk's own judgment agents/newsroom#wire-editor →

One assistant per solution page, each given the same facts that built the page and instructed to decline every other subject agents#ask-assistant →

Content + marketer agents drafting and shipping posts to blog.uzelhub.com on schedule agents/newsroom#content-agent →

Operator surfaces behind one capability URL: to-do list, draft review, publication queue, wire desk, agent roster, and a throwaway rebuild of the apex for walking copy see also: web.apex-sandbox

n8n as the automation substrate — workflows, webhook surface via ngrok, transactional data

HVAC business demo: n8n event generator, business dashboard, probabilistic meta-scheduler

Layered backups: nightly local staging + off-site B2 bundle, self-pruning, safe-reboot integrated

One shared Telegram channel for all agents — [AGENT]-prefixed, outbound-only, rate-capped with storm collapse, never blocks the caller agent-observability → see also: agent-platform.director.telegram

Reboot as orchestration: drain executions, snapshot, backup-or-abort, then reboot

Editorial site — 'a manifesto disguised as a portfolio'; literary register, process-as-product

Decision-free page production: one JSON record → generator → static site; template decided once, never hand-edit output uzelhub-site →

True stories from operating the systems, drafted from the lead queue, released on a drip and syndicated with canonicals see also: agent-platform.writer, web.publishing-drip

RAG pipeline: chunk-overlap ingestion, dual embeddings (Azure OpenAI / local HF fallback), FAISS + SQLite, metrics rag-framework →

Three domains read themselves every morning — film 06:00, semiconductors 06:45, weapons detection 07:30 — with a staleness alarm on a six-hour fuse

Zero-token CPU checks on every LLM output: evidence-in-source, orphan refs, zero-value docs ai-testing →

A cheaper understudy model silently re-runs primary-model jobs; disagreement is drift signal ai-testing →

Rolling 7-day gate pass-rate + quality-drift reporting; thresholds tuned on data ai-testing →

Scores never compare across a restart: history rows carry an epoch tag and a methodology-version stamp, so a change in code or source mix can't masquerade as a change in the world see also: predictor.two-lens-model

New domain onboarding from scaffold: ontology + feeds + prompts, no framework changes

Two clones, one data plane: dev pipeline feeds prod URL by symlink; deploy on merge

Self-contained web tests: Cytoscape mocked, real app JS, zero network ai-testing →

Specialist-facing taxonomy reference generated from domain profiles

Telegram-native command surface: long-poll listener, single allowed user, outbound sendMessage see also: ops.notify-telegram

An ambient 07:00 tick that writes the day's brief before anyone asks for it

Twelve thousand turns of agent session log, stripped to query/progress/answer and made queryable — the transcripts turned into a corpus see also: ops.root-consolidation

An append-only story queue the Scout may add to but never edit — 324 leads, forward-only status, mutations through one CLI verb

Every agent writes structured what-ran/what-it-decided/what-it-cost rows to one table — 1,401 decisions across twelve distinct actors agent-observability → see also: agent-platform.director, agent-platform.blog-pipeline

The platform checks its own pulse hourly and pages a human when it stops see also: ops.notify-telegram

Disk-bounded Docker: prune timers, log caps, builder GC on a 38GB box

A throwaway rebuild of the whole apex, every page in reading order with the data key behind each one, so copy can be walked and rewritten without touching production see also: web.voice-ledger

The crew introduced in plain language, each with what it does, a receipt, and the caveat that makes it honest agents →

Hand-tuned canvas animation on the editorial home (orbit glow, delivered-hold timing)

Cloud-quality embeddings with a local fallback — cost and availability as first-class design inputs

Retrieval quality measured, not vibed — per-stage metrics collected from day one rag-framework →

Landscape vs Movers: state-vs-change split that replaced 'trending' — each lens gets the view its physics demand see also: predictor.film-domain-rescue

Framework/domain separation enforced by a grep audit test — domains are config, never code about →

Ingest-only mode for short-form sources — count mentions, skip extraction, cut LLM cost

Domain selection by coverage fraction + calendar ground truth — where 25 docs/day is deep coverage

Ops calendar as code — .ics in the repo carrying drills, kickoffs, price-bump reminders

Monthly restore drill, every 3rd a full scratch-VPS rebuild from off-site alone — 'backups you haven't restored are a hypothesis'

_host/README.md as cross-project source of truth with dated corrections — drift treated as the primary failure mode see also: method.dated-amendments

The box's exposure written down and then audited against itself on a date — architecture doc, dated audit, and a running ledger of what was found

Two sites, same truths, different attention budgets — parity without apology, zero reuse

Editorial→marketing translation as a studied craft: concept inventory, carry?/drop? judgments, pattern appendix

Dropped-ideas tripwire: every Studio concept accounted for on marketing or explicitly left home see also: web.promotion-inventory

The box documents itself: dated prospecting survey → public man-page inventory, with a human-voice override layer, an id-stability contract, and now a written re-walk procedure inventory → see also: web.coverage-tripwire, web.survey-drift

Five-step register gradient: Home → cut sheets → field notes → inventory (man page) → Studio (literary); density earned per funnel step, deep-dive buttons are the stairs

The site's voice recorded as dictated rather than inferred — one entry per changed line, what was said, and the rule that falls out of it see also: web.apex-sandbox

One or two URLs a week, real dates never backdated, one promotion moment each — cadence as a guard against looking machine-made

Converse, Constrain, Construct, Curate — the working method behind everything here

100% human-AI collaboration, zero manual coding/tests, zero hardcoded domain logic — the maker's method as receipts about →

Docs never silently rewritten — corrections carry dates and keep the wrong claim visible ('corrected 2026-07-02: the previous text was wrong')

New domains/features declare their hypothesis and predicted scores before data exists — the validation instrument grades the prediction

Film 'failed' every fitness metric — until the Movers lens re-read failure as all-movement-no-landscape see also: predictor.two-lens-model

Smoke-test rows nearly fabricated rank deltas in production; 'a synthetic prior is worse than a cold start'

Rather than delete a backlog too old to score honestly, the deep pile was shelved — a new non-destructive status that keeps the documents queryable and the restart reversible see also: predictor.epoch-boundary

A branded persona built, shipped, and deliberately retired — what survives is the unbranded service

Agent asks to schedule a key-rotation reminder → refused as read-only → unreminded key expires → outage → diagnosing it wins the agent calendar-write authority → its first event is the reminder whose blocking caused the outage see also: agent-platform.director, agent-platform.ops-calendar

A metric meant to measure human-versus-agent agreement was quietly counting the agent's own marks as human ones — fixed by making the actor explicit at the point of marking see also: agent-platform.scout.leads-ledger

Off-site backups failed silently for 3 weeks (storage cap); every 'protected' bundle predated the fix that made dumps ride along notes/silent-backup-failure → see also: ops.notify-telegram, ops.restore-drill

The backup key documented as 'scoped, no delete' was account-wide with writeKeys — trust-but-verify via API, then rotate + Object Lock see also: ops.backup-system

Two installs merged into one user, and the abandoned one kept frozen copies of the same session ids — so the log ingest looked healthy while reading eighteen-day-old files see also: agent-platform.scout.session-ingest

The domain swapped under live traffic in one Caddy reload — staged noindexed, path-preserving redirects, no indexed preview possible by design

The 'content mine' was empty: four Studio pages were stubs, so marketing got grounded in the systems and now feeds the Studio

The self-updating inventory quietly went five weeks stale, still calling a daily-running agent 'designed-not-built' — and recovering how to refresh it meant reading a five-week-old session transcript, because the procedure was never written down see also: web.promotion-inventory, agent-platform.sysadmin

Retrospective precision@K / lead-time / graduation-rate instrument for 'is this predictive or noise?'

A second agent pointed outward, specified after an audit found something the daily health pass was not shaped to look for — its ledger is already seeded, it has simply never run agents#security → see also: agent-platform.sysadmin, ops.security-architecture