AI Research & Data Systems

A research team receives a steady stream of material from several sources: documents, pages, exports and messages forwarded by colleagues. Each item has to be classified against the team’s own scheme before it can be used, cited or searched.

Classification is judgement work. Two researchers given the same item would not always label it the same way, and the scheme itself had grown by accretion, with categories that overlapped and others that were rarely used.

Incoming material was classified by hand, inconsistently, and the backlog grew faster than the team could review it. Material waiting to be classified was invisible to the rest of the team, so the same source was sometimes researched twice and sometimes not at all.

Senior researchers spent their time on triage rather than analysis. Inconsistent labels made the record hard to trust, so people kept private copies and searched those instead. The backlog had no owner and no end date.

COST

Senior research hours spent on triage; duplicated research

RISK

Material missed or mislabelled; the record distrusted and bypassed

EVIDENCE

No record of who classified an item, when, or on what basis

The scheme, the review and the record belong to the research team. The system was shaped around that.

AUTHORITY

The agent proposes labels. A researcher accepts, edits or rejects every one. Nothing enters the record unreviewed.

SCOPE

Classification only, within the agreed scheme. The agent does not summarise, draft or search outside it.

PROVENANCE

Every item keeps its source, its arrival time and the reason given for its label.

DATA

Material stays in the team’s own storage. Nothing is used to train anything.

ADOPTION

One review queue inside the tools the team already uses. No new application to log into.

A shared inbox, a folder structure and a spreadsheet of labels maintained by whoever had time. The folders held the material, the spreadsheet held the classification, and neither knew about the other.

The head of research owns the system and the scheme. Every label in the record was accepted by a researcher, every acceptance and edit is logged, and the agent can be paused from the queue.

Intake is a single pipeline: material arrives by email, upload or API, is normalised to one record shape, hashed and deduplicated, and stored with its source. The agent runs on each new item against the agreed scheme and writes a proposal with a confidence and a short reason. Proposals land in a review queue; accepted labels go to the record, and every decision is logged with who made it. Before each release the agent is scored against an evaluation set built from the team’s own past decisions.

ASSESS

Intake mapped, scheme consolidated with the team, evaluation set assembled from past classifications, baseline defined

PILOT BUILD

Intake pipeline, record, agent, review queue, audit log, evaluation harness

EARLY DEPLOYMENT

In daily use; hardening and the measurement period in progress

Designed so that a wrong proposal costs a review, never a bad record.

EVALUATION

Scored against the team’s past decisions before every release; a drop blocks the release

CONFIDENCE

Low-confidence proposals are routed to review without a suggested label

QUEUE

If the agent is unavailable, items wait in the queue for manual classification; nothing is dropped

DEDUPE

Content hash and source match before an item is created; near-duplicates are linked, not merged

PAUSE

The head of research can pause the agent; the queue and record continue

A baseline was agreed before deployment. Figures will be published against it when the measurement period closes; nothing is reported before then.

PRIMARY

Time from arrival to classified record

SECONDARY

Researcher hours on triage; backlog age

QUALITY

Proportion of proposals accepted without edit; agreement with the evaluation set

BASELINE

The team’s records for the months before deployment

The system is in early deployment and its measurement period is not yet complete. Figures will be published here against the agreed baseline when it closes. Until then, only what is true by construction is reported.

• PRIMARY OUTCOME
  Measurement in progress
  Reported against the agreed baseline when the period closes.
• BY CONSTRUCTION
  Nothing enters the record unreviewed
  Every label has a proposer, a reviewer and a logged reason.
• BY CONSTRUCTION
  The backlog is visible
  Waiting items sit in one queue with an age, instead of in folders.

In daily use by the research team. Hardening continues: evaluation coverage is being widened and the scheme reviewed with the team. Results will be reviewed with the client when the measurement period closes.

PRODUCTION

Running in the client’s operation, measured, supported.

PILOT

One bounded system, live, with a baseline and an agreed measurement period.

EARLY DEPLOYMENT

In use, still being hardened; measurement period not yet complete.

TECHNICAL DEMONSTRATION

Built to prove an approach. Not operating a real workload.

CONFIDENTIAL CLIENT

Operating; client and figures withheld at their request.

FOUNDER-LED PRIOR WORK

Delivered by Stacx24 founders before the company; shown for method, not as company track record.