Taraz Systems

Enterprise Data Platform · AWS

Jarvis — real-time product defect intelligence at scale

A cloud-native, event-driven pipeline that ingests, deduplicates, and refines millions of product defect signals a day — replacing brittle batch jobs with a sub-second streaming design and becoming the authoritative source of truth for the whole organization.

Cloud ArchitectureCustom SoftwareAPI & Integrations

01The challenge

Product defect signals — missing attributes, item-level data quality issues — arrived from many upstream systems in high-throughput bursts. Brittle batch workflows meant downstream consumers acted on stale, duplicated, and conflicting data, and there was no single authoritative record per product entity.

02The strategy

Go fully event-driven and serverless. Kinesis Data Streams became the ingestion backbone to absorb bursts with sub-second latency; Lambda handled stream processing and refinement; and DynamoDB — with composite keys and conditional writes — guaranteed exactly-once, idempotent processing across retries and shard replays. Enriched records fan out over a second Kinesis stream, preserving a loosely-coupled pub-sub contract.

03The execution

Built the pipeline as discrete stages: ingest → normalize & enrich (Lambda) → dedup & state (DynamoDB) → reconcile & resolve schema drift (Lambda) → fan-out (Kinesis). Instrumented end to end with CloudWatch metrics, alarms, and structured logging, and tuned for horizontal scale, fault tolerance, and cost through auto-scaling Lambda concurrency and Kinesis shard tuning.

04The results

3M+
Defect events ingested at peak / hour
5
Pipeline stages, ingest to fan-out
100%
Exactly-once processing guarantee

See the system in motion

An interactive walkthrough of the architecture and data flow — hover, click, and drive the simulation.

Interactive system showcaseOpen full showcase

The pipeline, live

Raw defect events flow left to right — ingested, deduplicated, refined, and fanned out. Hover any node to inspect it and freeze its inbound traffic.

RawDeduplicatedRefinedRejected
  1. Upstream producer

    Catalog Feeds

    Bulk vendor and marketplace catalog feeds stream in item records with wildly varying completeness.

  2. Amazon Kinesis Data Streams

    Kinesis Ingest

    The ingestion backbone. Absorbs high-throughput bursts with sub-second latency and buffers them across shards for parallel consumption.

  3. AWS Lambda

    Processing λ

    Consumes Kinesis records, applies business rules, normalizes schemas, and enriches raw defect payloads into canonical records. Concurrency scales with shard load.

  4. Amazon DynamoDB

    DynamoDB

    Composite keys + conditional writes guarantee idempotent, exactly-once processing across retries and shard replays. Duplicates are rejected here.

  5. AWS Lambda

    Refinement λ

    Reconciles conflicting signals, resolves schema drift, and collapses many events into one authoritative record per product entity. Rejects malformed data.

  6. Amazon Kinesis Data Streams

    Kinesis Output

    Publishes enriched, high-fidelity records to a downstream stream, preserving a loosely-coupled pub-sub contract across the enterprise.

  7. Downstream consumer

    Analytics

    Data-quality dashboards read the canonical defect feed.

Drive the simulation. Push traffic, inject a bad record, or freeze the stream.

Events / sec

890

Duplicates blocked

0

P99 latency

47 ms

Downstream fan-out

3

Have a similar challenge?

Tell us where it hurts — we'll show you what's possible and what it costs.