The attribution layer of a production legal-intake platform: every paid click — Google, YouTube, Meta, TikTok — is traced from the ad through a three-domain landing topology, into the intake chatbot, and back out to the ad platforms as server-side conversion events. Plus a forensic methodology for the day the leads stop coming.
About this repo. Case study of a live client system; identifying details anonymized, code in
excerpts/sanitized and representative. Production source is private. All rights reserved — see the copyright notice below.
The firm runs paid acquisition across four channels, landing on three different domains (each channel maps to its own landing property). Before this work:
- Meta optimization was starved: browser pixels alone under-report (ad blockers, iOS), and there was no server-side signal at all.
- The dashboard couldn't answer "which channel produced this signed case?" — most leads arrived with no usable attribution.
- When lead volume dropped, nobody could tell whether the code, the tracking, or the campaigns had broken.
Every qualified lead fires a CAPI event from the backend, built to Meta's matching spec:
- Origin-routed pixel IDs. The three landing domains belong to different pixel properties. The pixel ID is selected per event from a Secrets-Manager map keyed by the lead's landing origin, with a safe default — one codebase, N properties, no cross-domain contamination.
- Browser + server dedup. The widget-side pixel and the server event share an
event_id, so Meta dedupes the pair instead of double-counting — the standard trap with hybrid pixel/CAPI setups. - Match-quality engineering. User data is normalized before SHA-256 hashing — phone to E.164, state to its 2-letter code, email lowercased — plus
external_id, client IP/UA, andevent_source_url, lifting Meta's event-match quality and therefore optimization performance. - Verified, not assumed. The most recent pipeline audit confirmed 100% CAPI delivery (42/42 events accepted) over the audit window.
A seven-bucket classifier stamps every lead at ingestion: Google, YouTube, Meta, TikTok, third-party lead vendor, CRM import, or direct/organic — from UTM parameters, landing origin, and ingestion path, with the ad name surfaced under each lead row in the operations dashboard so staff see which creative produced the person they're talking to.
Two honest lessons from production:
- The classifier is only as good as the tagging. A large share of paid traffic arrived with no UTM parameters at all, and one campaign shipped a literal
{adname}template token instead of the rendered value. Both are campaign-configuration bugs — the fix was UTM-tagged URLs handed to the marketing side, not more code guessing. - A bucket with zero traffic is a finding. The TikTok bucket exists and works; it read zero because the bio links were never tagged. The classifier's zeros located the gap.
Mid-year, paid leads abruptly stopped. The chatbot had just shipped a major update — so the update was the obvious suspect. We ran the investigation as a structured audit (parallel evidence tracks, then adversarial review of the conclusion) rather than a hunt for a comforting answer:
- Reconstructed the timeline from ingestion records: the drop aligned with the quarterly ad-campaign rollover on both Google and Meta — days before the widget update actually went live (which itself shipped three days later than planned; assumed dates were replaced with deployed-artifact evidence).
- Verified the pipeline end-to-end for the window: CAPI 42/42 accepted, intake alert emails 100% delivered, every CRM case in the window accounted for.
- Found the smoking gun in the client's own correspondence: the new quarter's ad had shipped with a malformed website URL (the bare UTM string, no destination).
Conclusion — with receipts: the code was exonerated; the campaign configuration was the cause. The write-up separated "what we verified" from "what we infer," and each claim was re-tested by adversarial review passes before it was reported to the client.
- Ad-platform integration done to spec: CAPI matching, hybrid dedup, multi-property routing.
- Attribution as a data-engineering problem: normalize → hash → classify → surface where decisions happen.
- Blameless-but-rigorous forensics: evidence over vibes, especially when the evidence clears your own code — a conclusion that's only credible because the methodology was adversarial.
| File | Pattern it demonstrates |
|---|---|
excerpts/capi_event.py |
Building a Meta CAPI event: normalization, hashing, origin-routed pixel, dedup event_id |
Built by Kamogelo Mahlasela and Masego Letsoko.
© 2026 Kamogelo Mahlasela and Masego Letsoko. All rights reserved.
This repository is published for viewing only, so prospective employers, clients, and collaborators can evaluate our work. No license is granted. Beyond viewing on GitHub (and the limited on-platform rights GitHub's Terms of Service provide), no part of this repository — text, architecture diagrams, or code excerpts — may be copied, reproduced, modified, distributed, or used to create derivative works without our prior written permission.