| --- |
| title: RevOps Agent |
| emoji: 🏢 |
| colorFrom: blue |
| colorTo: green |
| sdk: docker |
| pinned: false |
| --- |
| |
| # RevOps-Agent |
|
|
| **RevOps-Agent** is an OpenEnv-compliant reinforcement learning environment that simulates a real B2B Revenue Operations (RevOps) inbox. |
|
|
| In this environment, an AI agent receives inbound sales leads (from web forms, ads, etc.) and must perform full qualification and routing workflows that a trained human RevOps specialist would execute in a CRM. |
|
|
| ## Motivation & Domain |
| Every B2B company receives raw inbound leads that must be scored against their Ideal Customer Profile (ICP), checked against existing CRM data, and correctly prioritized. This operational layer is critical; bad routing results in lost sales equivalent to billions in wasted pipeline annually. |
| This environment bridges the gap by providing a meaningful, real-world task evaluation platform for tool-using LLMs. |
|
|
| ## Features |
|
|
| - **Trajectory-Shaped Reward:** Meaningful rewards are provided for interim operational steps like `enrich_lead` and `check_crm` rather than sparse binary rewards. |
| - **Strict Validations:** Penalizes actions that ignore basic protocols (e.g. scoring a lead before enriching it when necessary, or routing to a representative without checking the CRM). |
| - **Docker & Fast API Ready:** Exposes `/step`, `/reset`, `/state`, `/tasks`, `/grader`, and `/baseline`. Supported standard OpenEnv validation tool. |
|
|
| ## State & Spaces |
|
|
| ### Observation Space |
| The observation space (`RevOpsObservation`) is fully typed and represents what a RevOps analyst sees on screen: |
| - **`lead`**: Raw incoming lead information (name, email, message, etc.) |
| - **`enrichment`**: Firmographic data like company size, industry, revenue, region. |
| - **`crm`**: Current CRM status (existing accounts, active/closed opportunities for the lead). |
| - **`reps`**: A dictionary containing the roster of reps and their routing constraints. |
| - **`icp_criteria`**: Document explaining how to score and route leads based on segments and capacity. |
| - **`last_action_feedback`**: Human-readable string confirming the success or failure of the latest operation. |
| |
| ### Action Space |
| The action space (`RevOpsAction`) simulates RevOps workflows through discrete, typed operations: |
| - `enrich_lead` |
| - `check_crm` |
| - `update_lead_score` (requires `score`) |
| - `route_to_rep` (requires `rep_id`) |
| - `merge_with_account` (requires `account_id`) |
| - `flag_reengagement` (requires `opportunity_id`) |
| - `disqualify` |
| - `add_to_nurture`, `request_more_info`, `escalate_to_manager` |
| |
| ## Tasks |
| |
| Tasks natively support progressive evaluation. |
| |
| 1. **`task_easy`**: Direct Match Routing. Requires enrichment, simple scoring, and routing to the right Mid-Market rep. |
| 2. **`task_medium`**: Enrichment-Gated ICP Scoring. Lead has only an email; requires enrichment before scoring to avoid a penalty, and proper Enterprise assignment. |
| 3. **`task_hard`**: CRM Conflict. This lead corresponds to a closed-lost opportunity. The agent must merge accounts and re-assign to the original representative while dodging multiple concurrent junk leads. |
|
|
| ## Setup & Execution |
|
|
| ### Requirements |
| - Python 3.11+ |
| - [uv](https://docs.astral.sh/uv/) for fast dependency sync. |
|
|
| ### Local Installation |
| ```bash |
| # Optional: Setup openenv |
| uv pip install -e . |
| |
| # Validate environment compliance |
| uv run openenv validate |
| ``` |
|
|
| ### Running the API locally |
| You can rapidly boot the API for external agent inference scripts: |
| ```bash |
| uv run uvicorn server.app:app --host 0.0.0.0 --port 8000 |
| ``` |
|
|
| ### Docker |
| ```bash |
| docker build -t revops-env -f server/Dockerfile . |
| docker run -p 8000:8000 revops-env |
| ``` |
|
|
| ### Endpoints |
| The backend supports all standard OpenEnv mechanisms and adds: |
| - `POST /reset` - Resets and spawns the specific Task Scenario |
| - `POST /step` - Pushes a `RevOpsAction` to the state iteration |
| - `GET /tasks` - Lists tasks and JSON Action schemas |
| - `GET /grader` - Return final Grader Score [0.0 - 1.0] |
|
|
| ## Baseline Tests |
| Included is an automated baseline evaluator utilizing the OpenAI/NVIDIA NEMOTRON client. Provide `OPENAI_API_KEY` via `.env`. |
|
|
| ```bash |
| uv run python baseline.py |
| ``` |
| This triggers the three difficulties via a single agent loop and aggregates grader scores. |
|
|
