This write-up describes the system intent and engineering approach for an active build. It does not claim shipped functionality, measured performance, regulatory approval or customer outcomes that have not been published.
Design objectiveTurn a research brief into a reviewable evidence trail
A traceable path from input to outcome.
- 01Scope
Define the question, time window, languages, source set and inclusion rules.
- 02Collect
Use approved source adapters; retain retrieval time, source URL and connector status.
- 03Prepare
Normalize records, assess coverage, and flag duplicates or extraction gaps.
- 04Analyze
Retrieve relevant evidence and propose themes with source-linked spans.
- 05Review
Let a researcher inspect, correct and export the evidence-backed synthesis.
The research problem
Social research often begins with a broad question and ends with a collection of posts, comments and reports that are difficult to compare. A summary can hide the most important detail: which source supports a theme, how much of the available conversation was sampled, and what the analysis could not observe.
DEN Agentic AI is being built with a Canada-based startup as a social research intelligence tool. The technical goal is a research workflow that keeps collection scope, source provenance and analyst interpretation attached to every conclusion. It is not designed to treat a generated summary as a substitute for the underlying evidence.
A source-first technical flow
The proposed system separates acquisition, analysis and synthesis. Each stage records its inputs and completion state so a researcher can distinguish a real shift in conversation from a change in source coverage or connector behavior.
- Research brief: store the question, date range, language, source scope, inclusion rules and exclusions as a versioned study configuration.
- Source adapters: ingest only approved public or partner-provided material, preserving the original locator, retrieval timestamp, platform, language and connector outcome.
- Preparation: normalize text and metadata, record duplicate relationships, and surface missing pages, sampling limits and parsing failures instead of silently dropping them.
- Evidence index: make source records and relevant text spans retrievable through keyword and semantic search, with stable identifiers that survive re-indexing.
- Synthesis: have the research agent propose themes and claims that link back to retrieved evidence; route the result to a human reviewer before publication or export.
Keep the research agent inside a clear boundary
The agent should operate through a small, auditable set of research actions: search the configured corpus, open a source record, compare selected evidence, and draft a synthesis. It should not expand the source set, infer private attributes about contributors, or present a claim without an evidence path.
Each answer should carry source IDs, quoted or highlighted spans, date coverage and a confidence or limitation note. When the corpus is too small, the source is unavailable, or the evidence conflicts, the workflow should say so and ask for a narrower question or researcher review.
Data contracts, privacy and governance
A practical data model needs separate records for the study brief, source item, source snapshot, extracted text, theme, claim, evidence span, synthesis run and review event. Keeping these entities distinct allows a reviewer to see whether a conclusion came from a source, a transformation or a model-generated interpretation.
Connector-specific terms, rate limits, permitted retention and takedown requirements must be settled with the partner before collection is enabled. Access controls, retention periods and any handling of personal information should be configured to the approved study policy. This page describes the design direction; it does not claim a completed compliance review or a launched source integration.
How the system should be evaluated
Evaluation should test the complete research loop, not only a language model in isolation. A representative, reviewer-labelled corpus can establish baselines before the product is asked to support live research decisions.
- Measure source coverage, connector completeness, retrieval precision and recall, and whether citations actually support the attached claim.
- Compare theme consistency across repeated runs and time windows; stratify results by language, source type and topic where the data permits.
- Track analyst corrections, unresolved claims, review time, latency and cost per study, with failure cases retained for regression evaluation.
- Set release gates with DEN Agentic AI and the research users before publishing performance claims. No benchmark results are being asserted here.
Current build focus
The active engineering decisions are source permissions and connector boundaries, a stable evidence record, and the review surface that lets researchers move from a theme back to the underlying material. The first useful milestone is a small, bounded study that can be repeated and audited end to end.