Klaus Behnam Shad
AI Architectures · Research Infrastructure · Privacy by Design · Qualitative Data
Research
My work sits at the intersection of digital humanities, AI-assisted analysis and research infrastructure. I develop methods and tools for analysing qualitative data (interviews, narratives, documents) computationally, without replacing interpretation by automation.
Concretely, this produces rule-based analytical architectures, a multilingual evaluation harness and metadata models for interview-based research, among them DINOH, IMM-Core and tools for local, privacy-compliant processing. Traceability, reproducibility, data protection by design and the FAIR principles guide this work.
Theoretically, it rests on a sociocybernetic foundation: society as a recursive system of distinctions that reproduces itself through feedback, institutional routinisation and cognitive stabilisation. A methodological consequence follows from this: analytical categories are theoretically grounded and made explicit before coding begins. The method developed for this is Sequential-Configurational Narrative Analysis (SCNA).
Empirically, all of this is developed and tested on questions of social differentiation (belonging, exclusion, recognition) in fields such as migration, citizenship and oral history.
Research projects
Research software · exploratory
narrative-brain-laneA tested analysis pipeline for time-aligned audio, word and model features
Does a prosody band add to the prediction of high-gamma ECoG once word onset, mel spectrum, text predictability and Whisper features are already in the model?
A reproducible command-line Python pipeline computes fundamental frequency (F0), intensity, voicing, pre-word pauses, word onset, GPT-2 surprisal, mel spectrum and Whisper encoder features on a shared time base, each with provenance hashes. An audio gate admits only recordings whose checksum is on a local allow list. The pipeline is currently set up for the stimulus of OpenNeuro ds005574, which is used locally only.
Exploratory and descriptive. Across nine patients, the median added gain was about 0.0003 R² at reference-predictable electrodes and approximately zero across all electrodes. The planned residual-shift null failed a pre-specified synthetic check under simulated nonlinear redundancy, so no p-values are reported as evidence.
Exploratory analysis documented · over 270 automated tests
A local inspector prototype that shows how analytic tracks arise from a recording and how processing settings change them, tried out in a small usability study. A separate concept, Zweistimmig, would compare several readings of the same passage with listener markings; usage rights and ethics approval come first.