Dexa

• 

  Architecture

• 

  Application

• 

  Auto RL

Inspiration

• Supply chain in manufacturing is hard and personal.
• Teams juggle RFQs, certs, pricing windows, and lead time shifts.
• Scale breaks without adding a flaky middle layer.
• We wanted an agent that “is you” with vendors and customers.

What It Does

• Trains an RL model on your WhatsApp conversations to “be you”.
• Mirrors your tone, language, and nuances across contacts.
• Automates quotes, order entry, follow‑ups, and updates.
• Sends short, WhatsApp‑ready replies reliably.
• Auto‑retrains every 2 days to stay current.

## How We Built It

• Ingest: WhatsApp Business API (Meta).
• Orchestration: Mastra for multi‑agent flows.
• Models: Serverless RL on OpenPipe with Qwen3‑14B.
• Tracing/Evals: Weave (Weights & Biases).
• Infra: Google Cloud Run + Secret Manager + Logging.
• Web UI: minimal chat + “Be Like me” run button.

## Challenges We Ran Into

• Stable webhook auth and signature checks.
• Reliable short‑message splitting for WhatsApp.
• Safe RL loops with clear rewards and judges.
• Secrets, cold starts, and long‑running jobs on Cloud Run.

## Accomplishments That We’re Proud Of

• “Be Like me” one‑click RL trigger in production.
• Clean, fast web chat that also sends on WhatsApp.
• Two‑agent pipeline: reasoning → WhatsApp‑ready messages.
• Auto‑train every 2 days without manual ops.
• Clear logs and status tail for live runs.

## What We Learned

• Personal tone matters more than raw accuracy in ops.
• Simple JSON schemas tame LLM output for chat apps.
• Serverless RL works well with OpenPipe + Weave.
• Cloud Run is great for steady webhooks and bursty jobs.
• Keep it minimal: fewer moving parts, more reliability.

Built With

• google
• python
• weave