AIclicks: Product Engineer, AI Visibility Analytics

Built the tag system for prompts and topics: schema, API, tag chips in the prompts table, a dashboard tag filter, and bulk tag assignment at topic level. Added select-all and bulk actions on the prompts table with server-side selection semantics, so a customer can tag, pause or delete a whole filtered set at once.

Made the analytics views stateful: filters, tabs and sort live in the URL, prompt analytics sorting and a dynamic timeline follow one shared date range, and a project picker lets users who belong to several brands switch without losing context. A domain-view rollup aggregates brand pages across tracked domains.

Bulk topic delete went from one request per topic to one request for the whole set. On a 5,000-prompt fixture with a real Redis: 2,399 ms to 316 ms, 78 cache scans to 3, 26 webhooks to 1, 78 commits to 1.

• Bulk delete of 5,000 prompts: 2,399 ms to 316 ms and 78 commits to 1 (fixture, real Redis).
• Filters persist in the URL, so a view can be shared as a link.
• Prompt analytics sort and timeline follow one shared range state.

Built the generate-cluster-prompts pipeline in the FastAPI backend as eight stages. It embeds the brand's existing prompts, reduces them with PCA and clusters them with KMeans, picking k by a silhouette sweep from 5 to 40. Pools under 50 prompts skip clustering, and a clustering failure falls back to one bucket. A small model then labels each cluster as a topic.

A planning call to a larger model reads the labelled topics and returns a brief: saturated themes to avoid, forbidden phrases, and an instruction for the writer. A generation call writes 8 to 12 candidate prompts under geo-targeting rules. Candidates pass a cosine-similarity filter against the existing pool with two bands (0.70 and 0.92); the gray zone goes to an LLM judge, survivors are re-ranked by alignment minus avoidance similarity, and the top 10 are returned, with optional scoring.

In the recommendations service I extended the content-ideas path with the same idea at a different layer: a breadth-first search over co-cited URLs forms clusters, and each of the top 10 clusters gets one model call with its top 25 URLs to name the topic.

• Topic clustering replaced single-shot generation, so new prompts avoid saturated themes.
• Two cosine bands plus an LLM judge keep near-duplicates out without rejecting useful variants.
• Pipeline and cluster-stats behaviour covered by 10 tests I wrote; the content-ideas topic step has 17.

Designed the rollup tables and wrote the nightly explorer-preprocess worker as sole author: day planning, roll-day jobs, a claim guard with takeover, retry and heal paths, roll versioning and a late scan for stragglers. A gate serves a rollup only once it is complete for the day. The worker runs on a VM under Docker Compose.

On the API side I added the explorer metric cards, a citations API, a lean cited-sources endpoint, an influence route and an untracked-mentions route, backed by a GIN index on citations and a function-based query path. Compressing the cited-sources endpoint with brotli took the payload from 18.7 MB to 2.0 MB, a 9.34x cut measured in the real app stack.

This shipped as a first version. The rollup infrastructure needed a redesign before it could run across the full dataset, and that work was not finished when the engagement ended.

• Cited-sources payload 18.7 MB to 2.0 MB (9.34x) with brotli q4; 57 ms CPU against 238 ms for gzip.
• Reach queries after the citations GIN index: 106.8 s to 1.9 s on the heaviest site class, 153.9 s to 30.9 s over 14 queries with identical results (prod, one run each). Index 1.92 GiB on a 55 GiB table.
• Brand tab query 4.1 to 5.9 s down to under 0.1 s with a daily-brands rollup (prod mirror).
• Nightly worker: 90 of 90 days rolled, 0 errors, 163 to 190 s per day (prod, Oct 2026).

In the pipeline worker I added citation host matching and a test suite for the mention matcher before touching classification, fixed the stage-5 social host classifier, credited a brand's subdomains to the brand, and reworked mention credit accounting. On the backend I fixed the onboarding snapshot race so a new brand's first snapshot waits for its first analysis.

In the recommendations service I built the Get Mentioned path: citations enrichment with an on-demand variant, a ScrapingBee rate limiter, title-group jobs, scheduled mention fetches, Engage Online v2, the first phase of outreach feedback, and a reco-ops dashboard so operators can see job state without reading logs.

• Matcher review removed 1,814 wrong detections and kept all 5,674 correct ones (16,502 competitor rows against 38,804 page titles).
• Subdomain credit fix: 13.4% of tracked brands had uncredited subdomain mentions before it (prod read, checksum verified).
• Stage-5 brand-name fix recovered 15.8% more mentions for one customer, 174 snapshots rewritten and verified three ways.
• Operators see job state in a dashboard instead of logs.