Aimash — Google Ads work from Telegram
An Ads assistant for account managers: account data, reports and reviewed changes in chat. I built the workflow, approval controls and audit trail.
View the evidenceTimeline
2026
Role
Developer
4
Workflow stages
1
trusted approval per change
RU · EN
Chat languages
Engineering outcome
Read → review → decide → verify
Honest impact
An inspectable proposal, decision and audit workflow: the exact action and the human choice stay visible to the account manager.
Evidence from the work
What was built. What you can inspect.
- Delivered scope
- Telegram-to-Google Ads workflow: account reads, reports, reviewable proposals and controlled execution.
- What was checked
- Authentic redacted conversations dated 14, 16 and 30 September 2026, captured on 14 October. The budget example proposes 4 → 15 AUD and ends in cancellation.
- Evidence limits
- Historical captures, not a fresh live test. They do not prove an executed Ads write, exported XLSX contents or better client performance. The interactive website demo uses sample data.
Proposal and cancellation
A proposed daily budget change from 4 to 15 AUD, followed by cancellation. Historical conversation captured on 14 October 2026; no Ads write is demonstrated.
Interactive product demo
From a message to a verified change
Explore three everyday Google Ads workflows. Read a report, review a budget proposal, or cancel it before anything changes.
Interactive illustration with fictional accounts and sample figures. This page is not connected to Telegram or Google Ads; it does not show client performance results.
Aimash · Hermes
Demo · sample data
You
Increase the daily budget for Search · Demo from €50 to €60.
First, let the assistant prepare a proposal. You will be able to approve or cancel it.
Reports you can take away
Account analysis, period comparisons and spreadsheet exports for further work.
Ad copy ready for review
Generate search-ad headlines and descriptions, validate their limits and review a proposal before publishing.
Follow-up after a change
For supported changes, a read-only outcome checker compares later metrics. It does not autonomously adjust the account.
Under the hood: where the control lives
Telegram → Hermes Gateway → typed MCP tools → PostgreSQL proposal and trusted approval → Google Ads execution → audit and readback. Identity, freshness checks and one-time execution are enforced in Python, outside the model.
This case describes the implementation in the Aimash repository, reviewed in September 2026. A website demo is an illustration of that workflow, not a live production test or proof of revenue growth.
Have a workflow like this?
Start with one repetitive task: reporting, reviewing changes or connecting an internal system. We define the inputs, approval rules and a result you can verify before expanding the scope.
Discuss an automation pilotProblem
Routine Ads work spans account screens, spreadsheets and messages. Chat needs a clear boundary between a suggestion and permission to spend.
The Challenge
Show the target, current value and proposed change before execution, then record what happened.
The Solution
I built typed tools and a Python backend around Hermes, with stored proposals, trusted Telegram approval, one-time execution and Google Ads readback.
Technical Highlights
Typed MCP tools separate account reads, proposal preparation and approved execution.
Immutable proposals store the target account and the exact before-and-after values.
Approval is bound to trusted Telegram actor, chat and message context outside the model.
Freshness and one-time execution checks reject stale or repeated proposals.
Execution status comes from the audit trail and Google Ads readback.
Reports and supported follow-up checks can run without changing the account.
System Architecture
Python 3.12 · Hermes Agent + trusted Telegram transport
openai-codex (role-based model routing)
Google Ads API v25 (GAQL + guarded mutations)
Typed MCP tools for READ, proposals and approved execution
SQLAlchemy + Alembic + PostgreSQL
Pydantic tool schemas (function calling)
Hermes cron + read-only outcome scheduler
openpyxl + Google Sheets API
Fernet-encrypted secrets · audit log
Docker + GitHub Actions CI (ruff · mypy · pytest)
Core Technologies
Related Projects
Have a process that should run as a safe AI system?
I turn complex workflows into a scoped pilot with typed tools, approval gates and verifiable outcomes. You work directly with the engineer who builds and deploys it.