About

Goal: Assemble and harden our AI CX Agent so it reliably resolves real e-commerce tickets (chat/email, Shopify + Gorgias/Zendesk) with guardrails and monitoring. Scope: Wire LLM + tools (Shopify, Gorgias/Zendesk; optional: Klaviyo, Stripe, shipping). Implement intent router (orders, returns/refunds, shipping, product info, escalations). RAG over policies/FAQs; session memory with TTL. Guardrails: refund thresholds, PII redaction, escalation rules, rate limits/retries. APIs/Endpoints: chat endpoint, webhooks for email/SMS; auth + API key mgmt. Observability: structured logs, metrics, basic dashboards/alerts. Tests: golden conversation evals, integration tests for critical tools. Packaging & Deploy: Dockerized services, staging + prod parity. Docs & Handoff: README, .env.example, Postman/Insomnia collection, runbook, short Looms. Acceptance criteria (Definition of Done): First response, 2.5s avg; tool-using turns , 4s avg on staging. 85% pass rate on golden conversation set for top 8 intents. Clean human escalation in one click with full transcript/context pushed to helpdesk. Zero PII appears in logs (verified via redaction test). Reproducible deploy from a fresh checkout using documented steps. The skills required for your work: Back end: Node.js (TypeScript) or Python (FastAPI), REST/WebSockets, webhooks. LLM ops: Function/tool calling, prompt design, lightweight RAG (Pinecone/pgvector/Weaviate), evaluation harnesses. E-com/Helpdesk APIs: Shopify Admin, Gorgias/Zendesk (tickets, macros, tags), Klaviyo (profile lookups), Stripe (refunds/subs), ShipStation/AfterShip (tracking). Data & infra: Postgres, queues, Docker, environment/secrets management, basic AWS/GCP deploy. Reliability & security: Rate limiting, retries/idempotency, RBAC/ABAC for risky actions (refunds), PII redaction, audit logging. Testing & monitoring: Integration tests, synthetic conversations, OpenTelemetry/logging, metric dashboards. Good communication: Kickoff: You’ll review our code assets and propose a short architecture plan + milestone timeline before building. Weekly rhythm: 1x weekly 30–45 min check-in (live or async Loom) with a written status: done/next/risks/blockers. Daily async updates in Upwork messages (or Slack if preferred): commits, PR links, brief notes. Working artifacts shared early: Branches/PRs, staging URL, Postman collection, dashboard links. Expectations: Raise blockers within 24 hours. If estimates slip, propose a concrete replan (what changes, by how much, why). Availability: Overlap at least 2 hours/day with ET for questions and quick reviews. Details about how you or your team like to work Milestone-driven: We fund per milestone. Each milestone has an explicit checklist and demo to verify acceptance criteria. Source-of-truth repo: Git in our org; daily commits; PRs with short descriptions; no large “Friday dumps.” Small, verifiable steps: Ship a thin vertical slice early (chat → router → Shopify order status) before expanding intents. Security by default: Least-privilege API keys, secrets in a manager (not .env in repo), rotate at handoff. Operational quality: Logs/metrics/alerts exist before we call it “done.” If we can’t see it, we can’t trust it. Code style: Idiomatic TS/Python, linted/typed, clear module boundaries, comments on non-obvious logic. Include a short RUNBOOK.md with start/stop/redeploy and common troubleshooting. Documentation first launch: README with setup, .env.example, migration/seed scripts, and a 5–10 min Loom walkthrough. If the agent is delivered and everything works well, I will award bonuses with each client we obtain as long as the agent performs through launch. I will also extend an offer of maintenance work, that you would get first option on before I go in search of another engineer. In the SA/contract you will see this included.