MWDMotognon Wastalas d'Assise Dogbalou

Research / Educational Technology / AI & Data Science

STACK Analytics Assistant

A reusable deterministic analytics engine and web interface for trustworthy conversational analysis of STACK mathematics-assessment error exports.

All projects
ongoingPhD research project

Problem & context

Problem

Answer natural-language questions about STACK error data while keeping every numerical claim traceable to a deterministic, independently testable function and an explicit analytical population.

Intended users

Researchers and educators who need reproducible counts, distributions, item indices, comparisons and student filters from STACK exports.

Role and contributions

Analytical specification, deterministic backend, LLM routing and validation

Dataset / inputs

A standard nine-column STACK-AI error export, an optional question bank and an educator-supplied complete cohort size when whole-cohort facility or discrimination indices are requested.

Methods & architecture

The pipeline follows query to intent, deterministic function, structured result and grounded language. It separates the complete assessment cohort, error-analysis cohort and error-response dataset, and refuses whole-cohort metrics when their denominator is unavailable.

PythonpandasFastAPINext.jsOpenRouterCloudflare Workers

Implementation

  1. Validate reusable STACK export schemas
  2. Normalise free-text Newman error categories
  3. Provide deterministic statistics, rankings, profiles and compound filters
  4. Maintain multi-turn references without crossing uploaded datasets
  5. Isolate uploaded datasets and API keys by browser session
  6. Expose the engine through Python, CLI, FastAPI and a web interface

Current evidence

The current implementation verifies the project foundation below; broader outcome claims will be added after evaluation.

47 passedAutomated tests
3Analytical scopes
100 questionsAcceptance benchmark
DeterministicBackend mode

Limitations & next steps

Current status
The automated suite passes with deprecation warnings from the current Python test environment. The one-function-per-turn design keeps calculations traceable but requires a dedicated composite function for new multi-part question shapes. No public GitHub repository was supplied for this local project.

Publish a synchronised repository, refresh the Python test environment and evaluate routing and explanation quality separately from deterministic numerical correctness.

Attribution & licence

No licence file was verified in the local repository.