Build a Personalised Cold Outreach Pipeline: Site Research - Custom Video Pages - Instantly

What this is I run a small agency and I am building a repeatable outreach engine for one of my offers. Right now the research and personalisation is manual and it does not scale past about 20 prospects a day. I want it turned into a pipeline I can point at a raw list and let run. This is a scoped, paid test build on ONE sample list. If it lands well there is a much bigger follow-on (multi-campaign, scheduled, multi-client). The pipeline RAW LIST (CSV, ~200 businesses) | [1] CLEAN + VERIFY dedupe, strip role/generic addresses, MX check, email verification API, catch-all policy | [2] RESEARCH each business from its own website services, service areas, offers, contact paths, what is missing, what is claimed | [3] AI VISIBILITY SIGNAL ask an AI assistant a buyer-style question about their category + city, record whether the business is named, and who is named instead | [4] PERSONALISE (LLM, tightly constrained) one specific opening line + one P.S. per lead, grounded ONLY in facts pulled in steps 2 and 3, with a fallback tier when there is nothing good | [5] BUILD a personal landing page per lead one template, merge variables, their business name + logo or text header + video embed | [6] PUSH to Instantly via API campaign + custom variables per lead, scheduled sending, dry-run mode by default | [7] REPORT per-lead status table, what got personalised, what fell back, what was rejected and why What you are delivering A config-driven script or small repo (Python preferred) that runs the whole pipeline end to end from a single command, with a config file for the list, the niche, the campaign, and the API keys. List cleaning + verification step. Dedupe, remove role and generic addresses, MX check, run through an email verification API. You choose the verifier, tell me why, and state the catch-all policy. Target bounce rate under 2 percent. Per-business research from the business's own public website. Extract structured facts, not a blob of text. Handle the awkward cases: JavaScript-rendered sites, 403s, sites with almost no content, wrong or dead domains. AI visibility signal. For each business, run a buyer-style question about their category and city through an AI assistant API and record: was this business named, which businesses were named, and the source links cited. Store the raw response. Personalisation layer. One opening line and one P.S. per lead, generated from the structured facts only. Hard requirement: it must never invent a fact. I want to see your grounding approach and your fallback tiers when the research came back thin. Personal landing page per lead. One HTML template, merge variables, deployed as static pages (Cloudflare Pages, Netlify, or similar). Each page shows the business name, their logo (or a clean plain-text header when no logo is fetchable), and an embedded video placeholder. One unique URL per lead, passed into Instantly as a custom variable. Instantly integration. Push leads plus all custom variables into a named campaign via the Instantly API. Handle leads that already exist. Respect per-inbox daily caps. Dry-run is the default mode; live sending only behind an explicit flag. A QA report written per run: counts in and out at every stage, personalisation success versus fallback rate, rejects with reasons, and the API cost of the run. A README plus a short Loom walking me through a real run. Acceptance criteria (how the milestone gets released) I run the command on the sample list and it completes without me editing code. Dry-run mode makes zero sends and zero paid API calls beyond what is needed to build the preview. Running it twice does not duplicate leads in Instantly and does not double-send. Tell me your idempotency mechanism. I can open 10 random personalised lines and every single fact in them is traceable to that business's own site or to the recorded AI response. No hallucinated claims. The personalised landing page URLs all load and all show the correct business. The QA report numbers reconcile with what is actually in Instantly. Guardrails You will get a sample list and test inboxes only. No live client data, no client credentials. Nothing sends to a real recipient during this build. Test inboxes only. All API keys go in environment variables. Nothing hardcoded, nothing committed. Tell me the expected cost per 1,000 leads for the AI steps before you start building. Budget and timeline $350 fixed, one week. Milestones below. Everything stays on Upwork. To apply Skip the template pitch. Answer these four in your proposal: A prospect's site is JavaScript-rendered and returns almost nothing useful. Walk me through your fallback chain, and what the personalisation layer outputs when every tier comes back empty. How do you push a lead with custom variables into an Instantly campaign, and what happens on your second run when that lead is already there? Which email verifier would you use, what is your catch-all policy, and what bounce rate do you expect? Roughly what does the AI step cost per 1,000 leads, in dollars, with your assumed model and token counts? Link to something you have actually built if you have it. A repo or a Loom beats a testimonial.