Anjalikumari30/AI_Career_Agent_OpenEnv_Environment

AI Career Agent OpenEnv Environment

An OpenEnv-style benchmark for autonomous career agents. The environment models a real job-search workflow: finding relevant jobs, analyzing requirements, tailoring a resume, checking ATS fit, protecting authenticity, networking for referrals, applying selectively, and tracking outcomes.

Why This Environment Exists

Most resume projects stop at text generation. Real candidates need a fuller loop:

• —search selectively instead of mass applying
• —adapt resumes without inventing experience
• —improve ATS match without sounding machine-written
• —decide whether to apply now, revise more, or skip
• —network with employees and request referrals
• —track outcomes and interview probability

This environment turns that workflow into a reproducible benchmarkable task instead of a toy prompt.

Functional Requirements Coverage

1. Real-World Task Simulation

The environment simulates real career-search decisions rather than a game or puzzle. The core loop lives in career_agent_env/env.py, and the live workflow companion app lives in career_agent_env/live_app.py.

2. OpenEnv Specification Compliance

The repo now includes:

• —typed Pydantic Observation, Action, Reward, and EnvironmentState models in career_agent_env/models.py
• —reset() -> Observation
• —step(action) -> (Observation, Reward, Done, Info)
• —state() -> EnvironmentState
• —an openenv.yaml metadata file at the repo root
• —a FastAPI OpenEnv-compatible server at server/app.py

Validation command:

python -m openenv.cli validate . --verbose

3. Minimum of Three Tasks with Agent Graders

The environment provides three tasks with increasing difficulty and deterministic graders in career_agent_env/tasks.py.

4. Meaningful Reward Function

The reward model gives incremental feedback during the trajectory. It rewards ATS/authenticity improvement and useful actions, while penalizing spammy loops, low-quality submissions, and other undesirable behavior.

5. Baseline Inference Script

The baseline runner is scripts/run_baseline.py.

• —If OPENAI_API_KEY is set, it uses the OpenAI planner in career_agent_env/openai_policy.py
• —If HF_TOKEN is set, it uses the same OpenAI-compatible client against the Hugging Face router
• —Otherwise it falls back to the deterministic heuristic policy in career_agent_env/policies.py

OpenEnv Runtime

Metadata

The OpenEnv metadata file is openenv.yaml:

spec_version: 1
name: career_agent_env
type: space
runtime: fastapi
app: server.app:app
port: 8000

HTTP Endpoints

The FastAPI server exposes:

• —GET /health
• —GET /metadata
• —POST /reset
• —POST /step
• —GET /state
• —GET /docs
• —GET /web

Run locally:

uvicorn server.app:app --host 0.0.0.0 --port 8000

Observation Space

The Observation model includes:

• —candidate profile
• —current resume version
• —available job listings
• —active job id
• —current job analysis
• —ATS score
• —authenticity score
• —networking status
• —application history
• —interview probability estimate
• —remaining steps
• —applications remaining

Action Space

Supported actions:

• —SEARCH_JOBS
• —ANALYZE_JOB
• —MODIFY_RESUME
• —ADD_KEYWORDS
• —REMOVE_KEYWORDS
• —PERSONALIZE_RESUME
• —GENERATE_COVER_LETTER
• —CHECK_ATS_SCORE
• —APPLY_JOB
• —CONNECT_EMPLOYEE
• —REQUEST_REFERRAL
• —WAIT_FEEDBACK
• —SKIP_JOB

Reward Model

The typed Reward model returns:

• —value: scalar reward in [0.0, 1.0]
• —components: ATS improvement, authenticity improvement, bonus, and penalty

The main reward logic lives in career_agent_env/env.py.

Tasks and Difficulty Levels

1. smart_application (easy)

Apply once with ATS above threshold while keeping authenticity above the required floor.

2. strategic_campaign (medium)

Earn at least one interview with a limited number of applications.

3. career_optimization (hard)

Maximize interview rate across multiple companies while preserving authenticity.

Each task is scored by a deterministic grader returning a value from 0.0 to 1.0.

Data and Market Model

The environment uses dataset/Resume/Resume.csv to train a market model that powers:

• —category prediction
• —skill extraction
• —candidate profile construction
• —synthetic job generation
• —ATS alignment estimates

The training code lives in career_agent_env/dataset.py.

Live Workflow Companion App

The repo also includes a separate Streamlit app for a more human-facing workflow:

streamlit run career_agent_env/live_app.py

It can:

• —upload a real resume
• —infer a domain from the uploaded resume
• —fetch jobs from public Greenhouse and Lever boards
• —tailor resume and cover-letter drafts
• —score ATS and authenticity
• —export a tracker and application bundle

Setup

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Environment variables:

OPENAI_API_KEY=
OPENAI_BASE_URL=
OPENAI_MODEL=gpt-4.1-mini
HF_TOKEN=
HF_OPENAI_BASE_URL=https://router.huggingface.co/v1

Usage

Run the OpenEnv API Server

uvicorn server.app:app --host 0.0.0.0 --port 8000

Run the Baseline

python scripts/run_baseline.py

Train the RL Policy

python scripts/train_rl_agent.py

Run Tests

python -m unittest tests/test_env.py tests/test_live_workflow.py tests/test_server.py tests/test_openai_policy.py

Run the Dashboard

streamlit run career_agent_env/dashboard.py

Verified Baseline Performance

Verified locally on April 8, 2026 with the default seed (7) and heuristic fallback policy:

Aggregate baseline:

• —average_task_score: 0.7811
• —success_rate: 0.6667
• —policy: heuristic_fallback
• —seed: 7

Containerized Execution

Build and run the Docker image:

docker build -t career-agent-openenv .
docker run -p 8000:8000 career-agent-openenv

The container starts the FastAPI OpenEnv server and serves the lightweight web page at http://localhost:8000/web.

Hugging Face Spaces Readiness

This repository is prepared for a Docker-based Hugging Face Space:

• —Docker metadata is defined in the README front matter
• —the repo is tagged with openenv
• —the container entrypoint is the FastAPI server
• —the public web surface is served at /web

Repository Map

• —career_agent_env/env.py: environment loop
• —career_agent_env/models.py: typed models
• —career_agent_env/tasks.py: task registry and graders
• —career_agent_env/scoring.py: ATS and authenticity scoring
• —career_agent_env/dataset.py: dataset-backed market model
• —career_agent_env/policies.py: heuristic policy
• —career_agent_env/openai_policy.py: OpenAI planner
• —career_agent_env/live_workflow.py: live workflow orchestration
• —career_agent_env/live_app.py: Streamlit companion app
• —career_agent_env/dashboard.py: tracking dashboard
• —server/app.py: FastAPI OpenEnv server
• —scripts/run_baseline.py: baseline benchmark
• —scripts/train_rl_agent.py: RL trainer
• —openenv.yaml: OpenEnv metadata

Team

• —f-kansal-06
• —MikeanOP
• —Anjalikumari990