Book an intro with In Soo(a conversation, not a pitch)
  • Operational system
  • Data platform
  • Operational memory
  • Custom AI agent
  • Website
  • Software (SaaS)

Apshan

For Apshan, we designed a lakehouse that turns heterogeneous sources into structured, quality-controlled data available to agents. Custom CLIs run the routine flows autonomously; anomalies and rule changes remain subject to human approval.

Discuss a similar system ↗
01

The system delivered

Operations handled

  • Data ingestion
  • Labelling and enrichment
  • Data quality
  • Data orchestration
  • Memory retrieval
  • Agent data access

Architecture delivered

  1. 01
    Custom CLI
    Extraction and normalisation

    Specialised commands for each source and each stage of the pipeline.

  2. 02
    AWS S3
    Canonical object storage
  3. 03
    Apache Iceberg
    Versioned, interoperable tables
  4. 04
    Lakekeeper
    Apache Iceberg catalogue
  5. 05
    Qdrant
    Vector memory for agents
  6. 06
    Dagster
    Orchestration and observability
  7. 07
    AWS IAM
    Identities and least-privilege access
  8. 08
    A2A Protocol
    Interoperability between agents

Connected tools

  • AWS S3
  • AWS IAM
02

Operation and control

System capabilities

  • Autonomous extraction and normalisation

    Custom CLIs collect the sources, harmonise the formats and load the data into canonical storage.

    Runs automatically
  • Labelling and enrichment

    The pipelines apply Apshan's taxonomies and enrich each item before it becomes available.

    Runs automatically
  • Quality gates before publication

    Non-conforming batches are blocked automatically; only anomalies that need a decision are surfaced.

    Exception review only
  • Unified memory for agents

    Structured data and vector memory are available to authorised agents from a common foundation.

    Runs automatically
  • Agent-to-agent interoperability

    The A2A protocol exposes the capabilities and the memory to agents without duplicating logic or data.

    Runs automatically

Human controls

Approve new quality and labelling rules
Approval required
Handle data rejected by the controls
Exception review only
Change AWS IAM access rights
Approval required
Run conforming pipelines
Runs automatically
03

Operational memory

Structured operational knowledge

Extraction rules, taxonomies, labelling conventions, quality contracts, permissions and exception handling are structured in a single operational foundation.

Reusable corrections

A validated human correction becomes a normalisation rule, a labelling example or a control reusable in later runs. The system never changes its rules on its own.

04

Operational safeguards

Data qualityPublication blocked if controls fail
Every batch must meet the schema and quality contracts before entering the tables available to agents.
Access controlLeast-privilege rights per service
AWS IAM separates the identities and limits each component to the resources it needs.
ObservabilityObservable runs
The orchestration exposes pipeline state, failures and the steps that need intervention.
RecoveryVersioned data and tables
Canonical storage and Apache Iceberg snapshots make it possible to recover an earlier state of the data.
AuditabilityTraceable pipeline decisions
Validations, rejections and exception handling stay linked to the run that produced them.
05

Case study

The challenge

Apshan needed to collect and reconcile fashion data from multiple sources, with different formats, taxonomies and levels of quality. Manual processing could neither keep up with the volume nor produce a coherent memory for the products and the agents.

The system had to automate extraction and labelling without letting non-conforming data through, while keeping explicit access rights and human validation points for exceptions.

The solution

We delivered a unified data foundation: canonical storage on AWS S3, Apache Iceberg tables, a Lakekeeper catalogue and Qdrant vector memory. Dagster orchestrates the flows and makes every run observable.

  • Custom CLIs extract, normalise, label and enrich the data.
  • Quality gates block non-conforming batches from publication and surface the exceptions.
  • AWS IAM limits access per service; the A2A protocol makes the memory and capabilities available to authorised agents.
  • Apshan's bilingual website completes the system as the platform's public showcase.
06

Public interface, also delivered

Which part of your client work would you hand over first?

Book a call to map how that work moves through your agency today, agree the first workflow to build, and set what stays under your control.

See our workBook your free mapping call