Dr. Klaus Behnam Shad

AI Architectures · Research Infrastructure · Privacy by Design · Qualitative Data

Research Infrastructure

Research with interviews, testimony and institutional records runs into the same wall everywhere: the material is too sensitive to hand to a cloud service, and too unstructured to work through by hand at scale. The systems below are my answer to that. They are built to run inside an institution's own infrastructure, to stop rather than guess when something is unclear, and to leave a trail another researcher can follow.

Oral history infrastructure · Benchmark corpus · Seven languages

DINOH

Measures pipelines against a synthetic gold standard, so no real interview is exposed in testing.

Problem

There has been no reliable way to say whether an AI-assisted oral history pipeline actually works. Least of all across languages, and least of all without exposing real, sensitive interviews in order to test it.

Approach

DINOH provides a synthetic gold-standard corpus across seven languages, together with an evaluation suite, machine-readable schemas and WebVTT / OHMS export. Pipelines can be measured and compared before any real material is touched.

Status

MIT License · v1.0.0 · Archived on Zenodo, all versions: 10.5281/zenodo.21273366

Privacy gateway · Local-only · German, English, French

AegisQDA

Releases data downstream only after a human has reviewed and signed off.

Problem

Interview transcripts are full of names, places, employers and dates that make people indirectly identifiable. Under the GDPR that material cannot simply be handed to a cloud service for anonymisation, and automatic anonymisation on its own is not a sufficient safeguard either.

Approach

AegisQDA runs entirely on local infrastructure. Presidio-backed recognition of personal data is combined with mandatory human review; identifiers are replaced by typed, document-local surrogates and checked again in a second verification pass. Data is released to a downstream analysis tool only after explicit sign-off.

Status

MIT License · v0.1.0 (MVP) · Archived on Zenodo, all versions: 10.5281/zenodo.21854346

Qualitative data analysis · Local language models

DigQDA

Returns only what can be resolved back to a source location.

Problem

Language models produce coding that sounds convincing and cannot be checked. For qualitative research that is worse than no automation at all: an unverifiable claim about a source is not a finding.

Approach

DigQDA binds every output to its source. Versioned method contracts fix what is coded and how; prompts stay close to the source unit and return machine-readable JSON; segmentation is deterministic; quotes and locators are validated. A fail-closed conformance suite rejects results that do not resolve to the text.

Status

MIT License · pre-release (v0.4)

Metadata standard · Core profile · Cross-disciplinary

IMM-Core

Small enough to be adopted: thirteen fields, seven of them mandatory.

Problem

Interview data sits in incompatible silos. Every project invents its own fields, so material cannot be found, cited or reused across disciplines and repositories.

Approach

A deliberately minimal core of thirteen fields, seven of them mandatory, plus extensible implementation profiles for individual disciplines. Specified in DCTAP, with crosswalks to Dublin Core, schema.org/Dataset, REFI-QDA/QDPX and CMDI. Built for interoperability instead of lock-in.

Status

CC BY 4.0 · v1.0 · Archived on Zenodo, all versions: 10.5281/zenodo.20507328

Analytical framework · Rule-based · German, English, French

Relational Justice Analysis (RJA)

Fixes the coding rules before coding starts, so a second reader reaches the same result.

Problem

Qualitative coding is usually not reproducible. Categories emerge during the work and are rarely recorded in a form a second reader could apply to the same text and arrive at the same result.

Approach

Relational Justice Analysis makes the rules explicit up front: a rule-based, fully traceable annotation framework with indicators for German, English and French. Every annotation can be traced back to the rule that produced it.

Status

MIT License · v2.0.0 · Archived on Zenodo, all versions: 10.5281/zenodo.18640259

How these systems are built

Local-first

Processing happens inside the institution's own infrastructure. Sensitive data does not leave it. That is a property of the architecture, not of a policy.

Fail-closed

Where a system is unsure, it stops and asks. Silent guessing is the failure mode that makes automation unusable in this field.

Traceable

Every result resolves to a location in the source and to the rule or method contract that produced it.

Interoperable

Open standards and documented crosswalks instead of proprietary formats. The data has to outlive the tool.

Open

MIT and CC BY 4.0. Released versions are archived with persistent identifiers, so they can be cited, audited and built on.

These systems come out of concrete research settings: oral history, migration and citizenship research, institutional records. If you are working on something comparable, or need this kind of infrastructure for sensitive data of your own, I am glad to talk.

Get in touch