Backend Developer for Scoring Service

BACKEND DEVELOPER — SCORING ENGINE, CALCULATION SERVICE, AI GENERATION PIPELINE I need a backend developer to build a service that takes a completed 60-item assessment plus two pieces of personal data, computes scores and derived values deterministically, then drives a four-call AI pipeline that writes a long personalized report. The specification is finished. Roughly 80 pages. Every formula, weight, denominator, threshold, edge case and test case is already documented. You will not be guessing at requirements or waiting on me to decide something. Shortlisted candidates get the full spec under a simple NDA. WHAT YOU WILL BUILD 1. Scoring engine. 60 responses on a 0 to 4 scale, weighted across ten categories. Includes five reverse-scored items where a response of 0 yields the maximum contribution and a 4 yields zero. Normalization, two separate percentage layers, two eligibility gates, four validity rules, primary and secondary selection with documented tie-breaking. 2. Calculation engine. Derives seven values from a date of birth and a full name using a documented method. Fixed letter mapping, a digit-reduction routine with preserved exceptions, a three-branch rule for classifying one ambiguous letter per name-part, plus handling for accents, hyphens, apostrophes, multi-token surnames and non-Latin scripts. There is a built-in consistency check: two computed totals must sum exactly to a third. It runs in production, not only in tests. 3. Generation pipeline. Payload assembly to a documented schema, four sequential LLM calls (not one), deterministic sections assembled in code rather than generated, seventeen automated validation checks before anything ships, persistence, and call-level retry. Plus the glue: webhook in from HighLevel, callback out, idempotency so a retry never duplicates a record or a report. THE ARCHITECTURAL RULE The AI never calculates anything. Every number is computed in code, identically every time. The model receives calculated values and writes prose about them. Same inputs must always produce the same outputs. If your plan is to let the model do the arithmetic, we are not a fit. That is the one hard disqualifier. WHAT MAKES THIS HARDER THAN IT LOOKS The errors here are silent. If the reverse-scored items are implemented as normal items, nothing crashes and every result is quietly wrong forever. If a threshold reads the wrong percentage layer, nearly everyone lands in a fallback category and it looks like a product decision rather than a bug. If the letter-classification rule is applied inconsistently between two passes, the values are wrong and look completely plausible. That is why I care about tests more than speed. The spec includes reference profiles with expected outputs so this class of error gets caught before launch, not after thousands of people receive a wrong result. REQUIREMENTS - Python or Node, whichever you are genuinely strong in - REST APIs and webhooks - LLM API integration, built behind an interface so the provider can be swapped - Real unit testing. Not smoke tests. Tests asserting exact expected values. - Readable code another developer could pick up in six months - Secrets in a proper secret store - Raw birth dates and names must never enter the AI payload. Calculated values only. Nice to have: HighLevel webhook experience, or anything you have built where correctness mattered more than features. Volume is modest, hundreds to low thousands of runs per month. Do not over-engineer for scale. Spend the effort on correctness. MILESTONES 1. Scoring engine and test suite 2. Calculation engine and test suite 3. Generation pipeline 4. Validation, integration, documentation and handover I own all code, accounts and infrastructure at completion. I am interested in an ongoing maintenance relationship if the build goes well. SCREENING Before hiring I send a short paid test task, two hours at your rate, paid regardless of outcome. It is one page taken from the real spec, with worked examples so you can verify yourself as you go. Portfolios tell me little about how someone reads a specification, and that is the most important skill on this project. TO APPLY Skip the generic cover letter. Send me: 1. How would you verify that a reverse-scored questionnaire item was implemented correctly? Two or three sentences. This is the main thing I am reading for. 2. Python or Node, and why for this project. 3. One thing you have built that had to be correct rather than merely functional. Financial, billing, scientific, anything with a right answer. 4. Any question you have about the work. The best applications always contain one. I know exactly what I want built, it is documented, and I am ready to start.