A prop-firm review and complaint platform: traders compare prop-trading firms, file complaints and see how firms respond, with a trust score that reflects the outcomes.

Traders choosing a prop firm rely on reviews and word of mouth, and the firms with problems, such as late payouts or rules changed after the fact, are hard to tell apart from good ones. The platform gives complaints a structure: a trader files one against a firm, the firm can respond, moderators decide the outcome, and the result feeds a public trust score.
The hard requirements were fairness and resistance to abuse: a score that brigading can't move, a feed that surfaces live disputes rather than the oldest or loudest threads, and admin tools that are safe with several staff roles.
MongoDB through Mongoose holds firms, complaints, interactions and users. PostgreSQL has one job: the audit log, which is append-only and queried by time and actor.
A firm's score combines three weighted pillars: credibility, trading conditions and payouts. An upheld complaint deducts points, scaled down for firms that have paid out more in total, so one complaint moves a small firm more than a large one. Each deduction is recorded so it can be reversed or reduced if the decision changes.
Activity is a weighted sum. A firm's reply counts most, then staff replies, then traders reporting the same issue, then comments and upvotes. Trader signals are weighted by the voter's karma on a log scale and counted once per user. The hot score combines activity with time, Reddit-style; a second discovery score decays activity by age so new threads get seen.
Handlers push audit events onto a Redis list. A scheduled job pops them in batches and bulk-inserts them into Postgres, so logging never slows down or fails the user's request.

Complaints move through active, resolved, unresolved and rejected, with response deadlines, emails to the firm, and a scheduled job that resolves threads past their deadline.
Partial unique indexes allow one vote per user, or per guest session, for each interaction type.
Admin, moderator, editor and firm roles, checked by shared guard functions instead of per-route code.
Sliding-window limits per endpoint class, and uploads checked by file signature, size and dimensions before going to Cloudinary.
Firms propose edits to their own profile, limited to an allowlist of fields; staff approve or reject them.
Two databases to run is a real cost. The audit log is append-only and queried relationally by time, actor and action, which Postgres handles well and which keeps that load off the main database.
Queueing in Redis keeps requests fast, but an event can be lost if Redis fails before the batch job runs. For this product that was the accepted trade-off; a stricter audit requirement would need writes in the request's own transaction.