Micro Networks: a proof-of-concept for clawing back precise location
A POC model designed to extend Vicinity's existing location accuracy into the 80% of ad requests where users didn't share their coordinates.
- POCdelivered, not productionised
- Three rulesfrom Vicinity's actual data
- Integration scopedwith engineering, AdOps, Sales
I was Product Lead on Vicinity, a South African mobile advertising-technology business. The hero product was LEO, a proprietary ad server, and most of my time went to LEO platform work. The Micro Networks POC was one project within that — included here because it's the most self-contained artefact and a useful illustration of how I approach a product problem with data-side detail.
Vicinity's differentiator was real: when users shared their precise location, LEO served, attributed, and measured with a level of location accuracy competitors couldn't match. The opportunity sat in the other 80% of ad requests — the ones with only GeoIP, where Vicinity's edge was diluted because everyone had GeoIP. Micro Networks was a clawback design: infer the location of a WiFi hotspot from past user-shared coordinates clustered against its IP, so future requests from that hotspot could be served at the inferred location instead of GeoIP.
Design and prove the clawback. Use precise locations users had already shared to infer where their WiFi hotspots actually are; then design the routing in LEO that, on a no-share GeoIP request from a known hotspot, treats that request as an accurate WiFi-Home one. Build it as a POC in R, scope the production-integration and scaling paths.
- The Micro Networks insight: mobile users occasionally share their precise latitude and longitude. Group those points by IP address over time. If they cluster tightly enough, the IP is almost certainly a small WiFi hotspot — a home, small office, or café.
- Three pragmatic rules emerged from looking at Vicinity's actual data, not from a textbook clustering algorithm: cluster diameter ≤ 20m (anything bigger is a regional WiFi network, not a micro one); the modal lat/lon point per IP, not the mean (averaging puts the result in a road, dam, or empty field); and a minimum hit-count threshold (single-hit IPs are likely cellular, not WiFi).
- Integration design proposed for LEO: a new “Wifi - Home” targeting type for campaigns; the same WiFi type used by visit tracking to attribute in-store visits more reliably. Both designed to ride on the existing precise-location pipeline.
- Targets agreed for the project: convert 80% of GeoIP ad calls down to 30% by routing them through the inferred WiFi locations; lift WiFi click-through from 0.3% to 0.6%; roughly double campaign visit attribution by adding WiFi as a tracked location type. Targets, not measured outcomes.
- Scaling path articulated: the R prototype hit an in-memory ceiling around 50 million records. Production would have moved to on-demand AWS EC2 + Spark with sparklyr (the R—Spark bridge), provisioned by bash scripts that installed RStudio Server, Spark, and JAVA automatically. Designed, not deployed.
Micro Networks POC: cluster user-shared lat/lon by IP, apply three rules, output a WiFi-grade location per IP that the ad server would route to when users don't share.
Delivered a working proof-of-concept that validated the three rules against real Vicinity data. The integration design (a new WiFi-Home targeting type, plus the visit-tracking signal) was scoped with engineering, AdOps, and Sales. The scaling path from R-in-memory to AWS EC2 + Spark was articulated.
Production deployment was on the roadmap but did not happen during my time at Vicinity. The part I'm proudest of is the rules logic — it held up under scrutiny because it was honest about what the data could and couldn't tell you (excluding sprawling clusters, picking modal points to avoid roads and dams, gating on minimum hit counts), not because of a clever algorithmic step.
I was already Product Lead on the AdTech portfolio when an interest in data — and the gap in Vicinity's location-precision story — drew me into this project. Vicinity wanted to build a data-science function and the implicit assumption was I'd run it. Looking back, the company wasn't structured for a product manager to operate as an individual contributor on modelling from a product seat — and I wasn't experienced enough at the time to push back on that. Most of my time stayed on LEO platform work; Micro Networks is the one project I can show end-to-end because I did the data, the rules, the integration design, and the stakeholder scoping myself.