Technology

We are engineers first.

Most agencies operate someone else's software. We write our own. The site you are reading, the pixel that tracks it, and the measurement backend behind every client decision were all built here — and that changes what we are able to promise you.

What SingRank builds and runs in-house

SingRank runs a 74-tool measurement backend organised into ten nodes, driven by 79 documented workflows totalling 15,804 lines. Fifteen precomputed briefs regenerate on our own NAS — nightly for audits and ideas, every three hours for the content brief, using deterministic logic with no LLM call. This website is a Next.js static export on Cloudflare Pages: one codebase, three regional sites. Since 29 July 2026 we have used no Ahrefs and no Semrush; Google Search Console is our only keyword and demand source.

What we build with

Next.js and modern web. This site is a Next.js static export on Cloudflare Pages — one codebase producing three regional sites (Singapore, Indonesia, Malaysia) with edge functions for leads and content. Fast because it was engineered fast, not because a plugin promised it.

Python data pipelines. Our demand-mining pipeline crawls public sources politely — it reads robots.txt, honours crawl delay, and stores everything in a queryable database. What comes out becomes the raw material for content planning.

Local LLM and retrieval. The classification stage runs on a local model, so research text never leaves our machine. The backend pairs retrieval over our own indexed content with first-party performance data, which is why an answer can be traced back to a row rather than to a hunch.

Shopify, Wix and platform APIs. Fixes ship as code: Liquid theme work, structured-data injection, metafield updates, platform API calls. Implemented, not recommended in a slide.

Systems we built and run daily

• SingRank System — The 74-tool measurement backend behind every client decision, organised into ten nodes that run in one direction — from demand, through the archive, to ideas, faults, verdicts, evidence and results. Rankings, content performance, AI visibility, lead attribution and anomaly detection, regenerated on schedule.
• The skill library — 79 documented workflows, 15,804 lines, each one a written procedure rather than a habit in someone's head. It is why a task produces the same steps whoever runs it, and why we can show you the procedure behind a recommendation.
• Precomputed briefs — Fifteen documents rebuild on our NAS before the working day starts: the audit queue, the content brief, the idea bank, the competitor read. The content brief is pure deterministic logic — impressions weighted by position band and commercial intent — so it costs nothing to run and returns the same answer twice.
• SingRank Pixel — A first-party tracker with a detail most analytics miss: AI crawlers do not run JavaScript, so ours carries a no-JS component that records them anyway. It also identifies human visitors arriving from ChatGPT, Perplexity, Gemini and Copilot, and captures WhatsApp clicks and form submissions.

What we do when our own numbers are wrong

Measurement systems break quietly. Ours have. We write the failure into the operating manual rather than into a pitch deck, because a number nobody audits is not evidence — it is decoration. Three that are on the record:

A conversion rate above 100%. One report showed conversion rates that could not exist. The cause was a divisor fault: the numerator counted enquiries from every channel, the denominator only organic clicks from Google. Two different populations, so there was never a ceiling at 100%. The rule we wrote: do not rank work by that figure, and never quote it to a client until the denominator is fixed.

An attribution cutover we cannot undo. On 3 September 2026 our tracker began recording the page a visitor entered on. Before that it stored only the page where the form was pressed. Enquiries recorded before that date can never be re-attributed. So we do not compare per-page enquiries across that line, and a valid before-and-after only begins 28 days after it, with both periods on the same side.

A filter that silently did nothing. One tool accepted a date-range argument and ignored it: fourteen days and ninety days returned an identical figure. We traced it to a row cap shared across the whole roster, documented it, and named the correct source to use instead. Empty or flat output now triggers a health check before any conclusion, because a dead collector looks exactly like an honest zero.

AI, used the way an engineer uses it

We use AI where it is strong — research, classification, drafting, extraction — and keep people where judgement lives: strategy, accuracy, and what gets published in a client's name.

We run models locally when privacy demands it, retrieval when an answer must come from real data, and we log what AI systems actually do to our own site instead of guessing. That is also why our GEO work is measured rather than theatrical.

Common questions

Why this matters to you

An agency that builds its own instruments cannot hide behind a tool's estimate. When we show you a number we can show you the system that produced it — and when that system is wrong, you will hear it from us first.

Talk to the team that builds it