Description
Summary I run Cinch, a direct-to-consumer camping gear brand (Shopify). I need a developer to build a scheduled, unattended service that (1) triages our customer-service Gmail inbox by applying labels, and (2) writes ready-to-send reply drafts for selected products using the Anthropic Claude API, grounded in a knowledge base I maintain in Notion. Nothing is ever auto-sent — I review and send everything in Gmail. Important: the design work is already done. This system has been prototyped end to end and validated on the live inbox. The attached brief (v2.1) is a complete specification: architecture, label taxonomy, routing and precedence rules, drafting rules, voice spec, guardrails, extensibility model, operational lessons from the prototype, and testable acceptance criteria. This is an integration + deployment project, not a discovery project. Please read the brief before proposing — proposals that don't engage with it will be declined. That said, the brief is not scripture: reasoned challenges to the design are welcome and count in your favour — see §17 of the brief for the ground rules (engage with the documented lessons; show your alternative still meets the acceptance criteria and guardrails). Stack (open to your proposal): Python or Node; Gmail API (server-side OAuth); Shopify Admin API (read-only); Notion API (the knowledge base lives in Notion); Anthropic Claude API with prompt caching, with all LLM calls behind a provider-agnostic adapter (LiteLLM or equivalent); scheduled serverless or small always-on host; secrets management; basic logging/alerting. What I'm asking for in your proposal: Answers to the three screening questions below (required — proposals without them will be declined). Your scoping questions after reading the brief — and any design challenges, per §17. A proposed architecture and phased plan. A fixed quote for v1 against the acceptance criteria in §16 of the brief, plus your position on the open decisions in §18. Screening questions (answer all three in your proposal): This must run unattended on a server. How would you authenticate to Gmail for that, and what are the trade-offs between the options? Section 11 of the brief describes an incident where a Gmail search silently returned nothing and the run nearly reported success having done no work. How would you prevent that class of bug? The pipeline runs three times a day and must be safe to re-run at any time. How would you guarantee re-runs never duplicate work (e.g. never draft the same email twice)? Working arrangements (non-negotiable): All accounts and assets in my name: Google Cloud project, hosting, Anthropic API key, Notion integration, and GitHub repo are created under my accounts, with you added as a collaborator. Shopify access is a read-only token I create. The system never sends email and never writes to Shopify — hard guardrails, see §9 of the brief. Milestone-based fixed price: Milestone 1 is a paid written scoping response (settling the brief's open decisions and raising any §17 challenges); subsequent milestones are triage live, then drafting live, each accepted against the brief's acceptance criteria; final milestone includes a handover runbook (setup, secrets, how to redeploy, how to diagnose a failed run). About you: you've shipped production integrations against at least two of: Gmail API, Shopify Admin API, Notion API, Anthropic/OpenAI APIs. You write software that fails loudly and logs well. You're comfortable being held to written acceptance criteria. Timezone: I'm in the UK; async is fine, some overlap for a weekly check-in appreciated.