Personal AI Operations Agent

Self-hosted multi-agent system. One Telegram chat for the entire business and personal productivity across 21 functional modules.

A tool-use orchestrator with ~95 LLM tools, hybrid RAG over corporate documents, an auto-extracting knowledge graph, financial automation at 97% accuracy, 8 background daemons. Single-VM deploy at €5/mo. Full source code, no vendor lock-in.

I want this agent Send a Brief

Solo entrepreneur|Self-hosted|6 weeks (Phase 0 → 26)|Telegram + Web Dashboard

21: Functional modules
~95: LLM tools
49: DB migrations
14: External integrations
8: Background daemons
289: Documents indexed
97%: Finance auto-categorization
€5/mo: Hetzner VM

Operating a one-person business from memory and 17 browser tabs

Each pair below is a real scenario that used to cost hours every week.

Without

Notes scattered across Apple Notes, Google Drive drafts, email saved, Telegram saved, paper. ~30% of ideas get lost.

With

Single inbox capture: voice / photo / text into Telegram → auto-classify → routed to the right store (task / fact / contact / event / commitment / journal).

Without

Finances done by hand: export CSV from each bank, categorize 200+ rows / mo manually, project month-end on a piece of paper.

With

Auto-sync with 5 bank accounts. 97% LLM auto-categorize, recurring-merchant rules, EOM forecast with category breakdown, bulk-undo for mistakes.

Without

Overdue tasks pile up in layers. Deadlines lost in the stream. Promises to people forgotten — trust erodes.

With

Smart task lifecycle: due_at coercion (odd hours → end-of-day), 2-tier dedup (exact + fuzzy), auto-snooze stale > 7 days with one nudge, commitments auto-expire 24 h after due.

Without

“Where was that pricing email from X / decision in some Drive doc / note about Y” — 20–40 minute hunt every time.

With

Hybrid RAG search (BM25 + vector cosine) over 289 unified documents (email + Drive + voice). Answer in 1–2 sec with source citation [src: email «Pricing inquiry from X»].

Without

Voice messages garbled in noisy environments (car, street), Ukrainian / mixed terms misheard.

With

Production voice pipeline: gpt-4o-transcribe + ffmpeg pre-process (highpass 80 Hz, FFT-denoise, EBU R128 loudnorm) + auto-detect lang (RU/UK/EN) + entity-name prompt context.

Without

LLM occasionally invents actions (“I deleted task X”) without an actual tool call.

With

Two-layer anti-hallucination: action-verb-name match in system prompt + supervisor judge (gpt-4o) verifies every reply against tool_results. Hallucination → reply replaced with system fallback. Logged in /quality dashboard.

How the agent works: the request pipeline

Every tool-use call passes through 8 layers — from security guard to supervisor judge. Hallucinations are caught before the reply reaches the user.

Quick check

in-process regex denylist

Validate

agent_security: input sanity, prompt injection guard

Routing

agent_brain LLM router → tier (cheap / mid / premium)

Cost gate

downgrade @80 % budget, hard-block @95 %

Recall

knowledge graph + session summary injection

Tool loop

Claude Sonnet 4.6, max 6 iterations, ~95 tools

Verify

supervisor judge (gpt-4o) anti-hallucination

auto-split > 4096 chars, strip markdown for TG

Phase 20 architectural refactor: 11 previously-separate sub-agents (capture, scheduler, inbox, drive, tasks, surface, finance, wellness, journal, subscriptions, viz) moved into backend/skills/*.py — they now run in-process in the orchestrator. Container count: 21 → 10. Zero HTTP latency for hot tool calls.

One Telegram chat. 21 functional modules. Zero context switching.

Compact catalog: modules grouped by purpose. Full tool list and ~95 endpoint details live in the project's docs.

Personal productivity

Capture

Universal inbox: voice / photo / text → auto-classify

Tasks

Smart due-coercion, 2-tier dedup, auto-snooze stale, escalation

Routines

Recurring cadence (cron), streak tracking, building/steady/wobbly trend analytics

Life

Goals + dreams + achievements + profile across 12 life areas

Reminders

TG push from notifications table with auto-dedupe and priority filtering

Wellness

Habit logging + daily state (mood / energy / focus) + summary analytics

Business operations

Finance

Monobank auto-sync (5 accounts), 97% LLM auto-categorize, EOM forecast, bulk-undo

Subscriptions

Auto-detect from email, weekly renewal alerts at +7 days, recurring tracking

Inbox

Gmail every 15 min, per-email LLM classification, auto-task for high+requires_action

Drive

GDrive metadata + textual ingest (Google Docs / text / PDF) into RAG store

Calendar

Two-way Google Calendar sync, conflict detection, plan-for-window queries

Commitments

Promise tracking, owe / awaiting direction, auto-expire 24 h after due

Knowledge & memory

Documents (RAG)

Universal documents table: 289 docs / 441 vector chunks, hybrid BM25 + vector search

Memory graph

entities + facts (subject-predicate-object), 4-step entity resolve, /memory react-flow UI

Conversations

Long-session summary daemon: > 50 msg → gpt-4o-mini summary injected next message

Library / PKB

Legacy knowledge base with daily / weekly / monthly automatic summaries

Privacy, system & ops

Journal

AES-GCM 256 column-level encryption, key never logged, never plaintext in DB

Surface

Morning Brief / Evening Recap / Weekly Review (cron)

Approvals queue

Human-in-the-loop for destructive / high-cost / bulk > 10 ops, 24 h auto-expire

Quality dashboard

verdict per LLM call (ok / suspicious / hallucination), 24 h hallucination rate

Cost tracking

Single source of truth for LLM / embedding / voice ops, EOM projection, hard caps

LLM routing: cheap tasks run on cheap models

agent_brain classifies the request and picks a tier. Premium hits Opus 4.7, fallback chain swaps to Haiku/Gemini when the primary is down.

Tier	Model	Use case
Premium	Claude Opus 4.7	Complex planning queries
Default	Claude Sonnet 4.6	Orchestrator main loop (1 h prompt cache, −47 % input)
Mid	GPT-4o	Supervisor anti-hallucination judge
Cheap	GPT-4o-mini	Categorization, summarization, entity extraction
Embedding	text-embedding-3-small	1536-dim vectors for hybrid RAG
Voice STT	gpt-4o-transcribe	Auto-detect language with entity-name prompt
Fallback	Claude Haiku 4.5 + Gemini 2.0 Flash	Provider outage chain

What's under the hood

AI core

Claude Sonnet 4.6 / Opus 4.7 · GPT-4o + mini · Haiku 4.5 · Gemini 2.0 Flash · text-embedding-3-small

Voice & input

gpt-4o-transcribe · ffmpeg pipeline (highpass / FFT-denoise / EBU R128) · Telegram Bot API + MTProto Pyrogram · Web Speech

Storage

PostgreSQL 16 · pgvector (1536-dim ivfflat) · pg_trgm · 49 migrations · 18 roles RLS · AES-GCM 256

Frontend

Vite + React 19 + TypeScript · TanStack Query · react-router v7 · react-flow · Recharts · framer-motion

Infrastructure

Hetzner CX22 VM (€4.99/mo) · Ubuntu 24.04 LTS · Docker Compose (10 containers) · Tailscale Funnel · least-privilege service user

Backend

Python 3.12 · FastAPI 0.115 · uvicorn · asyncpg · pydantic v2 · OpenAI / Anthropic / Google GenAI SDKs

What an agent like this costs

Build is paid once, runtime lives at $65–$130/mo. For comparison: ChatGPT Team is $25/user with no custom tools, no bank integrations, no automation logic.

1 — Agent Development

$6,000

one-time setup

Tool-use orchestrator + 14 in-process skills + Docker Compose deploy + 30+ dashboard pages + 49 DB migrations with RLS + 18 per-agent roles + 2-layer anti-hallucination + RAG ingest pipeline + voice quality pipeline + full docs (ARCHITECTURE / INFRASTRUCTURE / SECURITY / PHASES). 1 month support after deploy.

2 — LLM Operating Cost

~$50 – $100

per month

Claude Sonnet 4.6 main loop ~$2–5/day (1 h prompt caching = −47 % on input), embedding pipeline ~$0.95/day, GPT-4o supervisor judge ~$10–20/mo, gpt-4o-mini categorization ~$5/mo. Hard caps: $5/day, $100/mo, $0.05/task.

3 — Optional add-ons

$5 – $30

per month each

ElevenLabs voice TTS for TG voice replies $5–$22/mo · Vapi outbound voice calls $0.07/min · Twilio SMS notifications $0.01/msg · Hetzner Storage Box off-site backup €4/mo.

4 — Infrastructure

€5

per month

Hetzner Cloud CX22 VM (2 vCPU / 4 GB RAM / 40 GB SSD / 20 TB egress). Tailscale Free Personal $0. GitHub private repo $0. Total runtime ~$65 – $130 / mo (= $780 – $1 560 / year) with infra + LLM + add-ons.

What actually changed

30-minute hunt for «that email/doc about X» → 2-second RAG search with citation

0 missed subscription renewals — all in weekly brief 7 days ahead

97 % auto-categorized finance (1073 / 1102 tx) — was 0 %

5–15 % more accurate voice transcription in noisy environments (ffmpeg pre-process)

Container count: 21 → 10 after Phase 20 merge — −65 % RAM vs initial architecture

289 documents indexed, 441 vector chunks — searchable knowledge corpus

8 background daemons run unattended — no manual triggers required

Cost transparency — daily / monthly / per-agent / per-operation breakdown, hard caps. No surprises.

Want this agent? 6 weeks from zero to production.

Self-hosted multi-agent system — one Telegram interface for the entire business. Single source of truth for tasks / finance / knowledge / contacts. LLM-powered automation with anti-hallucination guarantees. Full source code + docs. $5 VM, no vendor lock-in.

Book a call (30 min) Send a Brief

The architecture grows for years: realtime voice WebSocket, Whisper.cpp self-hosted, GitHub / Notion integrations, multi-tenant variant for team deployments.