Case study: Marketplace intelligence

An audit engine for marketplace listings

A working audit experience that turns product listings, reviews, and competitor evidence into clear recommendations.

Try the live demo30 seconds · representative sample data · nothing leaves your browser

Client work2025–2026In production

Project details

Year: 2025–2026
Role: Product · engineering
Status: In production
Stack: Cloudflare Workers · tRPC · Durable Objects · D1 · Drizzle · Next.js · R2
Relationship: Selected delivery work; the public case uses a functional project name.
Market: Marketplace brands; client identity is not disclosed publicly.

Try it

Run the representative audit.

Pick a sample product to replay the workflow and inspect the stages, evidence panels, and recommendation format.

Interactive demo, using representative data in your browser

Scraping listing
Mining reviews
Mapping competitors
Scoring & recommendations

Pick a product and run the audit. The pipeline replays the workflow on representative data. Nothing is generated live; this illustrates the system's output shape, end to end.

The problem

A marketplace audit that keeps the evidence beside every recommendation.

Product teams were comparing listings, reviews, keywords, and competitors by hand. The work was slow to repeat and difficult to review once the market changed.

ZAWISH made a browser audit experience and analysis workflow that collects representative inputs, scores clear gaps, and produces traceable recommendations.

The system

How it fits together.

Scroll sideways to explore the enlarged map.

Read the components

Listing data and review snapshots feed the analysis engine. Job state and storage support processing; the outputs are an evidence-backed score and recommendations.

Listing scrape: structured data
Review dataset: async snapshot
Job state: Durable Object
Analysis engine: matrices · gaps
Storage: D1 · R2
Audit score: evidence-backed
Recommendations: listing rewrites

simplified public map of implemented components and boundaries

The build

Keep the evidence connected to the decision.

Reviews behind a login wall

Amazon gates deep review pages. Instead of fighting it per-request, the system triggers an asynchronous review-dataset snapshot early in the run and processes it in parallel with the competitor scan, polling schedules, error fallbacks, and cost settlement handled by the worker.

Dataset API · async polling

Six listings, one comparable matrix

Bullets, taglines, descriptors, image counts, and video metadata from six competitors are normalized into structured copy and visual matrices , so a gap is a cell you can point at, not an impression.

matrix services · LLM analysis

Recommendations that cite their evidence

The rewrite engine aggregates review pain points and keyword gaps, then drafts listing copy where every suggestion links back to the cluster or competitor row that justifies it. No "best practices" filler.

recommendation engine

Serverless, because audits are bursty

The whole backend runs on edge workers with a per-project state room coordinating job progress. Capacity can scale with audit demand while job state remains resumable and inspectable.

Workers · Durable Objects · D1

Implementation evidence

Source snapshotsraw listing and review inputs retained for traceability

Comparison matrixcompetitor, copy, visual, and review signals aligned for review

Evidence linksrecommendations remain connected to their supporting observations

Constraint

The output needed to preserve source evidence while several analysis jobs completed asynchronously.

Tradeoff

The system favors traceable snapshots and resumable jobs over a fast but opaque single prompt.

The interactive audit is a representative browser demonstration. It illustrates the workflow without exposing client data or claiming commercial uplift.

For your business

Discuss a related system.

The public demo uses representative data and replays the workflow in the browser. It is not connected to a client account.

Start a project More work

Next caseA 3D placement pipeline from Blender to render farm →