AI Pipeline — Chat History to Knowledge Base + Qualification Flow

AI Pipeline Developer — Chat History to Knowledge Base + Qualification Flow (Ruby or Python) ABOUT US Leadster is a Brazilian SaaS company working with lead generation and conversational marketing. We run a WhatsApp customer service platform used by hundreds of businesses. THE PROJECT When a customer connects their WhatsApp number to our platform, we get access to their conversation history. Today, turning that history into a working AI agent is manual work: someone reads the conversations, writes a knowledge base, builds a qualification flow, and tests it. We want to automate that end to end. You will build a pipeline that receives a conversation history and produces three things: 1. A knowledge base document, built from the topics found in the history. 2. A qualification flow (question script), based on how that company actually qualifies leads. 3. A PDF report with 10 simulated conversations, scored and reviewed by an LLM judge, with concrete improvement suggestions. The output goes to the customer for review. Nothing is published automatically. SCOPE — 6 STEPS Complexity is shown as filled circles out of 5. 1. Consume conversation history from our API — ●○○○○ 2a. Generate knowledge base document from history — ●●●○○ 2b. Generate qualification matrix and flow script — ●●●●○ 3. Test environment: simulate 10 distinct conversations — ●●●●○ 4. LLM judge: score, classify, generate PDF report — ●●●○○ 5. Deliver the PDF to our endpoint — ●○○○○ Full technical specification is shared with shortlisted candidates after an NDA. STACK - Language: Ruby or Python, your choice. - LLM: OpenRouter, keys provided by us. No model restriction — pick per step based on task complexity. - Hosting: your proposal. The full run is long, so it must be asynchronous, not request-response. - Delivery: GitHub repo under our organization, CI configured, tests, documentation. WHAT WE PROVIDE - Base prompts for the LLM steps. These are a starting point, not final — you are expected to iterate. Large changes should be discussed first, with the reasoning and the expected outcome. - A sample JSON of the message format, so you can start on day one. - Mocks for any endpoint of ours that is not ready yet, so your work is never blocked. - OpenRouter keys. Some of our endpoints are still being built internally. We work around that with mocks — your final documentation should show how to swap them for the real integration. WHAT WE ARE LOOKING FOR - Solid experience building LLM pipelines in production, not just prompt experiments. - Comfortable with prompt iteration as an engineering task: measuring output quality, not guessing. - Clean, tested, documented code — this will be maintained by our team afterwards. - Direct communication and willingness to push back when something in the spec does not make sense. Portuguese is a plus but not required. Our team works in English for this project. PROPOSAL — TWO ROUNDS Round 1 — your proposal. Include: 1. Your hourly rate. 2. A short description of a similar LLM pipeline you have shipped, and what went wrong with it. 3. Which model or models you would pick for step 2b and why. 4. How you would validate that the generated flow is actually good, given there is no fixed rubric. Round 2 — estimate. Shortlisted candidates sign an NDA and receive the full specification. You then send the estimated hours per step, which we turn into fixed-price milestones. Do not send a total price in round 1 — you do not have enough information yet. TERMS - Contract type: fixed price per milestone, following the step breakdown above. - Expected start: immediately. - Code ownership transfers to Leadster. NDA required before the full specification is shared.