| title | EETRE - Enterprise Email Triage & Response Environment |
|---|---|
| emoji | 📧 |
| colorFrom | blue |
| colorTo | green |
| sdk | docker |
| pinned | true |
Inbox chaos? Same.
EETRE is an OpenEnv environment that trains an LLM agent to handle enterprise email like a real ops teammate.
It does 4 practical actions:
replyescalatearchivemark_spam
And yes, escalation can send real Slack alerts.
I was missing important emails because of volume.
Promotions, spam, notifications, and real incidents were all mixed together.
So I built EETRE to answer one question: Can an agent make correct workflow decisions in a real inbox, not just classify text?
- Theme #1: Multi-Agent Interactions
- Reasoning Agent -> Decision Agent -> Auditor Agent
- Theme #3: World Modeling (Professional Tasks)
- Real mailbox actions via IMAP/SMTP + Slack escalation webhook
| Item | Details |
|---|---|
| Agent sees | sender, subject, body, priority |
| Agent can do | reply, escalate, archive, mark_spam |
| Reward based on | action correctness, response quality, efficiency |
| Modes | simulated curriculum + live mailbox |
- Reasoning Agent reads the email and finds intent/risk.
- Decision Agent picks one action.
- Auditor Agent blocks unsafe choices (example: replying to suspicious mail).
- Environment executes action and returns reward.
Source files: plots/eetre_training_results.png, plots/eetre_reward_curve.png.
Embedded images load from GitHub raw so they work in the GitHub UI and in the Space README without storing PNGs in the Space Git repo (HF rejects large/binary PNG pushes unless you use Git Xet). Push origin first so these URLs resolve, then push the Space using SPACE_PUSH.md (Xet or the no-binary scripts/push_hf_lite.sh).
| Metric | Value |
|---|---|
| SFT loss reduction | 37.5 -> 18.1 (51%) |
| GRPO reward improvement | 0.608 -> 0.833 (37%) |
| Training steps | 96 |
| Avg trained reward | 0.836 |
Problem → system → proof → impact in five steps.
Mixed inbox with spam, notifications, and critical emails.
Example of a production issue email requiring immediate attention.
The agent escalates critical incidents to Slack with structured context.
Meaningful replies instead of generic auto-responses when a reply is the right action.
Third-party tool run on this repo for a quick sanity check—not part of EETRE’s training stack.
Context: Nitish Kulkarni (ex–Google ML/AI, building NextToken) emailed in April 2026 with a free preview flow: exploratory analysis, model training, and interactive dashboards in a notebook-style agent. He linked a pre-populated prompt to run on openenv-email-triage-benchmark (the screenshot above is from that run). EETRE itself stays OpenEnv + TRL + your Space; NextToken is optional external validation and a nice “someone else drove the repo in a UI” datapoint for judges.
| Step | What it signals |
|---|---|
| Inbox | Problem |
| Incident | Trigger |
| Slack | Action |
| Reply | Intelligence |
| Benchmark | Credibility |
curl -X POST https://Vetri17-openenv-email-triage-benchmark.hf.space/reset \
-H "Content-Type: application/json" \
-d '{"task_id":"medium"}'- HF Space: https://huggingface.co/spaces/Vetri17/openenv-email-triage-benchmark
- Push Space from Git:
SPACE_PUSH.md(Git Xet or no-binary scriptscripts/push_hf_lite.sh) - Colab (training): https://colab.research.google.com/github/Vetri1706/openenv-email-triage-benchmark/blob/main/notebooks/eetre_training.ipynb
- Training script (local / TRL):
eetre_grpo_final.py— Python script that runs GRPO (Group Relative Policy Optimization via Hugging Face TRL) against your live HF Space/reset+/stepAPI; produces the metrics above and can save a fine-tuned model. - GitHub: https://github.com/Vetri1706/openenv-email-triage-benchmark
- Blog / Video (<2 min): https://youtu.be/ndwF3Rp_f2Q
EETRE turns inbox chaos into actionable workflows — detecting, deciding, and executing in real time like an ops teammate.