SubTag

Founders know Reddit works — threads rank on Google, buyers read them, and AI assistants cite them. What kills the channel is operations: which of ten thousand threads are worth showing up in, what to say that a subreddit will accept, and how to do it consistently without a team.

The agency behind SubTag had solved this manually, placement by placement, for hundreds of brands. My job was to turn that human operation into software.

One engineer — me — owning everything from schema design to checkout flows. That constraint shaped the architecture: boring, proven pieces (Next.js, Supabase/PostgreSQL) arranged so one person can operate them, with background jobs doing the heavy lifting instead of services that need babysitting.

The core of the product is a scoring engine that answers 'which threads deserve your budget?' It ingests Reddit threads for a workspace's keywords, then ranks them on two axes: live search traffic (is this thread actually ranking and pulling visitors?) and buyer intent (are people in it deciding what to buy, or just chatting?).

Getting this right was a data-pipeline and ranking problem, not an LLM problem: joining traffic signals to threads, deduplicating, scoring intent, and keeping the list fresh as threads rise and die. The result is a ranked opportunity list a customer can act on directly — the feature everything else hangs off.

Comment drafts are generated with a multi-step chain rather than a single call: draft → critique → revise. Each workspace carries its own brand-voice context that is injected into the pipeline, and every AI draft goes through human review before anything is published — the system is built to assist an operation with a quality bar, not to spray content.

Most of a real SaaS is not the AI. SubTag has isolated client workspaces (users → workspaces → projects with per-workspace roles), a credits-based billing system across four plan tiers, schedulers and background jobs that must not double-execute, real-time analytics, and a public API. Each of those is a correctness problem — billing especially. Building them solo is the strongest engineering exercise I've had.

SubTag is in beta, evolving weekly with early users — I deliberately don't quote usage numbers yet. The next engineering investment is a formal eval harness for draft quality: regression prompts per subreddit style, an LLM-judge calibrated against human review decisions, and cost/latency budgets per pipeline stage.