Agent Systems Workspace

ML/backend workspace for Beton's agentic analytics systems. This repo contains the backend agent logic that powers Beton product experiences around signal discovery, warehouse reasoning, and evidence-backed analysis workflows.

flowchart LR
    UserPrompt[User prompt] --> ProjectSelect{Project}
    ProjectSelect --> Upsell[Upsell Ranker]
    ProjectSelect --> Telegram[Telegram Analytics]
    Upsell --> UpsellFlow[Website or account context -> Beton Inspector-backed warehouse analysis -> signal search]
    Telegram --> TelegramFlow[Chat target -> schema bootstrap -> SQL-backed insight loop]
    UpsellFlow --> Outputs[Reports, signals, ranked opportunities]
    TelegramFlow --> Outputs
    Scripts[Optional local scripts] -. probes .-> Outputs

Loading

The repo keeps multiple agent lineages side by side so architectural changes stay explicit. The latest versions of both projects are optimized for iterative analysis and are currently most reliable when deployed with Beton Inspector in front of protected systems. A standalone mode that connects directly to systems such as PostHog, Attio, and similar integrations is planned but not implemented yet.

Product Context

This repository is the inspector-ml-backend for Beton. It is not the full product surface by itself.

• Product website: www.getbeton.ai
• Pricing and hosted trial: www.getbeton.ai/pricing
• Integrations catalog: www.getbeton.ai/integrations
• Demo stand / product-facing repo: https://github.com/getbeton/inspector/tree/staging

Use those references to understand the product this backend connects to:

• getbeton.ai is the public product entry point for Beton.
• The getbeton/inspector staging branch is the demo stand and product-facing layer that sits in front of this ML/backend repo.
• This repository provides the agent workflows and warehouse-analysis backend that the Beton product layer can call into.

Projects

• projects/upsell_ranker contains the upsell opportunity lineage.
• projects/telegram_analytics contains the Telegram chat analytics lineage.
• scripts/ contains optional local utilities for one-off probing rather than the main agent runtime.

Repository Layout

projects/
  upsell_ranker/
    versions/
      v0.0.0/
      v0.0.1/
      v0.0.2/
  telegram_analytics/
    versions/
      v0.0.1/
scripts/

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

Fill .env with only the variables needed for the version you want to run. .env.example contains placeholders only and should not be treated as working credentials.

How To Test

Use adk web as the primary local workflow. It starts the ADK server with a web UI, which is the intended path for interactive testing, prompt iteration, and debugging.

Upsell ranker:

adk web /home/user/upsale-agent/projects/upsell_ranker/versions/v0.0.2

Telegram analytics:

adk web /home/user/upsale-agent/projects/telegram_analytics/versions/v0.0.1

Keep adk api_server for production-style or external API integration scenarios:

adk api_server --host 0.0.0.0 --port 8000 /home/user/upsale-agent/projects/upsell_ranker/versions/v0.0.2
adk api_server --host 0.0.0.0 --port 8000 /home/user/upsale-agent/projects/telegram_analytics/versions/v0.0.1

Legacy baseline smoke check:

python -c "import sys; sys.path.append('/home/user/upsale-agent/projects/upsell_ranker/versions/v0.0.0'); from scoring import rank_accounts; print(rank_accounts(top_n=5))"

Environment

Core variables used in the repo:

GOOGLE_API_KEY=
AGENTOPS_API_KEY=

ATTIO_TOKEN=
ATTIO_LIMIT=200
ATTIO_MAX_PAGES=10
ATTIO_CACHE_TTL_SEC=21600

INSPECTOR_URL=
AGENT_SECRET=
VERCEL_PROTECTION=

UPSELL_AGENT_MODEL=gemini-3-flash-preview
DWH_ANALYST_MODEL=gemini-3-flash-preview
SIGNAL_AGENT_MODEL=gemini-3-flash-preview
SIGNAL_REVIEWER_MODEL=gemini-3-flash-preview

SIGNAL_AGENT_MAX_QUERIES=6
SIGNAL_AGENT_MAX_ROWS=200
SIGNAL_AGENT_MAX_ITERATIONS=8
SIGNAL_AGENT_TARGET_PROMOTED=3
SIGNAL_AGENT_MAX_FAILURES=5
SIGNAL_AGENT_MIN_ITERATIONS_BEFORE_EXIT=3

TELEGRAM_ANALYTICS_MODEL=gemini-3-flash-preview
TELEGRAM_ANALYTICS_REVIEWER_MODEL=gemini-3-flash-preview
TELEGRAM_DB_HOST=127.0.0.1
TELEGRAM_DB_PORT=5432
TELEGRAM_DB_NAME=
TELEGRAM_DB_USER=
TELEGRAM_DB_PASSWORD=

Open-Source Notes

• This repository is licensed under AGPLv3. See LICENSE.
• Beton Inspector is currently the security-focused product path for protected warehouse access in the newest agent versions.
• Direct standalone integration flags for systems such as PostHog and Attio are planned.
• Recent experiments show that explicitly defining the target metric for an agent is the strongest lever on analysis quality.

Notes

• Generated artifacts, caches, logs, and local report outputs are intentionally ignored.
• Version-specific details live in each project and version README.
• Public docs use "warehouse" in place of internal shorthand such as "DWH" on first mention.

---

About

This backend is part of Beton, open-source revenue intelligence. The agent it powers is described at getbeton.ai, and its native data integrations are documented at getbeton.ai/integrations.

Related Beton open-source tools:

• Inspector — the front-end agent + workspace this backend serves
• DryFit — synthetic analytics datasets for agent benchmarking
• openclaw-gtm-skills — company research pipeline for OpenClaw

Browse all open-source tools by Beton →