Local models / Task pipelines / Bulk processing

Orchestrate local models. Decompose complex tasks.

AI & Knowledge develops the processing layer for research: explicit task pipelines, suitable model roles and AI-assisted classification of texts and images.

AI Scaffolding is a local prototype for task normalization, semantic routing, specialized model roles and reviewed outputs. It provides the basis for further decomposition of composite tasks and integration with Nea Polis.

Capabilities

Areas of work

Local Model Orchestration

Assign model roles and resource requirements to subtasks. Local processing is a deliberate architectural choice; model responses still require review.

Task Decomposition

Break composite tasks into verifiable steps with defined inputs, outputs and dependencies.

Text and Image Classification

Classifiers support bulk processing of texts and images. Taxonomies, ambiguous cases and manual review are part of the intended workflow.

Knowledge & Research

Source references, structured results and review connect AI processing with the research record. A model response does not automatically become canonical evidence.

From a Composite Task to a Verifiable Output

Prototype documented through read-only inspection: task normalization, semantic routing, specialized workers, critic and validators. General automatic task decomposition into a dependency graph and direct Nea Polis integration remain development objectives.

01
Normalization
Convert the raw request into a technical task specification and incorporate workspace context.
02
Task Pipeline
Process tasks.md sequentially; manage task IDs, states, bounded retries and resumption of interrupted tasks.
03
Routing
Assign workflow and model roles, including General, Editor, Retrieval, Coder, Vision and text classification.
04
Worker
Execute roles through the local Ollama connection. Classifiers process individual objects or batches specified by a manifest.
05
Validation
Critic and deterministic validators examine delivery contracts, artifacts, taxonomies and change boundaries for each workflow; code paths use a shadow workspace.
06
Result
RunResult separates success, delivery_mode, answer_text, artifact_path and feedback. Batch traces support traceability.

Bulk processing: text & images

Text Classification

Implemented in code: taxonomy-based text classification, geo taxonomies with aliases and structured object outputs; individual and batch pathways.

Batch manifest, source references and taxonomy contract are checked; structured results feed subsequent processing steps.

Image Classification

Implemented in code: single-image analysis and vision_batch with a freshly generated manifest, isolated prompt context for each image and explicit error results.

Closed categories or hierarchical taxonomies; trace artifacts and batch-coverage checks keep processing auditable.

From processing to research

Development objective: controlled decomposition of composite research requests into subtasks and dependencies, with results handed to Nea Polis. The existing sequential Task Runner does not prove a general DAG scheduler. Reviewed model results do not automatically become canonical evidence.

Do not replace original data with model results.
Keep model/configuration versions and task references traceable.
Distinguish unclassifiable and erroneous outputs from valid categories.
Do not equate model confidence with empirically calibrated reliability.
Bulk processing requires verifiable failures and review, not just overall success.
Local-first does not mean fully offline: retrieval workflows may access network sources.
A shadow workspace and policy checks are safeguards, not a demonstrated operating-system sandbox.

Source revision ai-agent-framework, commit f8849c981a34d40905b33e312cf070a4f8b044e7, read-only inspected on September 2, 2026: README, orchestrator, task runner and selected regression tests. No models launched and no full runtime test.

Source repository

Applied systems

Project archive

Agentic SystemsPrototype