Five things an AI-run company actually shipped.

Every morning someone opened the iREV CRM, pulled reports per client, eyeballed funnels, screenshotted spreadsheets, and pasted them into a digest. Hours of copy-paste. Worse, when a lead API silently broke, nobody knew until a client noticed — sometimes a full day later.

An automated pipeline pulls the daily iREV reports on a schedule, analyzes each funnel, and assembles a per-client "report card" written for the client. Delivery runs through Nina into Telegram. The same pipeline watches the lead and API layer in real time and alerts the moment an endpoint errors.

The morning copy-paste-and-screenshot ritual is gone. Client-ready report cards land before the team is at their desks, every day. Manual lead-checking hours go to zero, and a broken API surfaces in minutes instead of a lost day of revenue.

Operations don't keep office hours. Reports need to go out, alerts need to fire, errors need a response at 3am and on Sunday. Hiring around the clock is impossible for a lean team — so things get missed in the gaps.

Nina, an always-on AI employee that lives in Telegram over MTProto, so she works like a real account, not a bolted-on bot. She delivers the daily digests, runs live alert checks every five minutes, and nudges the team the instant something errors.

The work that used to require a person watching dashboards now has a worker assigned to it permanently — no shifts, no gaps. Five-minute monitoring coverage, 24/7. The same role a human ops hire would do, at a flat subscription instead of a salary.

A busy operator drowning in email, tasks scattered across apps, calendar and reminders living in someone's head. Important things slip not because nobody cares, but because nothing connects the inbox to the action.

Muninn, a multi-user assistant on Telegram with persistent memory. It reads incoming email and turns the actionable ones into real Google Tasks automatically — invoices, lawyer and accountant emails, school notices. It manages the calendar, fires reminders that actually exist, and runs research on demand. It runs on a Claude subscription, not metered per-token billing, so cost stays stable — and it's a closed, supervised system, not an open agent framework.

Email stops being a backlog and becomes a task queue that fills itself. Reminders the system claims are set are actually set — "done means done" is enforced, not assumed. One assistant covers triage, tasks, calendar, and research.

For a huge number of SMBs, the entire customer relationship lives in Telegram DMs — deals, support, leads, follow-ups, all scattered across threads with no pipeline. Buying a traditional CRM means dragging the team off the platform they already use. They won't, so they don't.

CRgraM, a Telegram-native CRM built on MTProto. Instead of forcing customers into a separate system, it turns the Telegram chats they already run into a structured CRM — contacts, pipeline, and follow-up tracking inside the tool the business already lives in.

SMBs get CRM discipline without the migration that kills CRM adoption. The pipeline lives where the conversations already happen, so it actually gets used — the rarest and most valuable property a CRM can have.

Every AI agent with memory has the same quiet bug: it stores the wrong things as truth. A user asks "Do I have a meeting Friday?" and the agent files "user has a meeting Friday" as a fact. Commands, hypotheticals, sarcasm — all written to memory as gospel. In one production system, this corruption hit health-record data.

speech-act-memory-gate — an open-source (Apache-2.0) ingest gate that sits in front of any AI memory layer and decides what's actually a fact versus a question, command, or hypothetical. A drop-in: a few columns and one function call. It pairs with a confidence model so facts carry doubt, gain certainty when restated, and decay when stale.

On SpeechActMemBench, the gate moved fact-storage precision from a baseline of 35.2% to 100% — it stops storing the garbage that poisons agent memory while keeping the real facts. It's public on GitHub, with the benchmark you can run yourself.