The seven working groups
Last updated 31 July 2026
Seven project working groups are launching together. Each one exists to produce something specific and finishable — not to discuss a topic. Every group is looking for two co-leads: a domain lead who owns the substance of the deliverable, and a program lead who runs the machine (cadence, progress notes, recruiting, unblocking).
You can rank the groups you want to join, and nominate yourself as a co-lead. Both are open now. A group without both seats confirmed stays forming; nothing stalls the others.
Each group below is a summary. The full charter for every group is published — mission, deliverable, scope, milestones, and what it needs — and each summary links to it.
For how groups are chartered and how they operate, see the RFP.
1. Ansari 4
Build the next generation of Ansari on the MultiSage expert-panel architecture: several frontier models consulted in parallel, combined with Ansari’s Islamic knowledge tools — Qur’an search, hadith retrieval across the major collections, prayer and date awareness, scholarly citations — with a facilitator model deciding when to consult the panel and synthesizing cited answers, including comparative fiqh perspectives across madhabs.
First deliverable: a deployed Ansari 4 (alpha → member beta → public), evaluated against Ansari 3 before public release.
Grounding: scholarly review is essential here, not decorative. Panel synthesis must not launder weak opinions into confident answers.
Works with: Islamic Benchmarking provides the evaluation gate; Standards for Islamic AI would make Ansari 4 the reference implementation.
2. Islam, AI & Education
Build an adaptive learning platform for Islamic education of 13–18 year olds.
First deliverable, staged: a curriculum and pedagogy specification grounded in the tradition with a scholarly review loop; then a prototype covering one subject with a pilot cohort; then iteration toward a deployable platform.
In scope from day one: 13–18 means minors. Privacy, safeguarding, and parental consent belong in the charter’s scope section at the start, not discovered later.
3. Non-profit AI Acceleration
Help Muslim non-profits adopt AI well. The founder’s sketch has three stages — case studies, transformation services, and an AI-first operating system for non-profits — and stage one is the charterable unit.
First deliverable: a published case-study collection of Muslim non-profits that have successfully implemented AI (target 8–12 organizations, interviews plus write-ups). It is finishable, needs no code, and its findings de-risk the later stages: what do non-profits actually adopt, and where does it stick? Stages two and three become successor charters informed by what stage one learns.
4. Post-training Frontier Open-Weight Models for AI Sovereignty
Post-train frontier open-weight models to hold Islamic values, and release the recipes and weights openly — so the community owns frontier models aligned with its values, rather than renting them.
Which model: base-model selection is part of the work. Inkling (Thinking Machines) is the leading candidate but only an example — Nemotron is also on the list. Inkling’s case is empirical: mid-pack out of the box, but the best guided ceiling of any model tested, with near-zero steadfastness drop under pressure. The values fit in a page of guidance and the model holds them, which is the ideal profile for baking them in through post-training.
First deliverable: a frontier open-weight model post-trained to hold Islamic values, evaluated on JaleesBench against its base model and the guided ceiling, with the recipe and weights diff released openly.
Starts with: defining the input to the post-training. A values spec grounded in Qur’an, Sunnah and maqasid is the starting hypothesis, not a foregone conclusion — and whatever form it takes, it is a religious document. Scholarly review here is load-bearing, not decorative.
Works with: the tightest loop in the pipeline. This group trains, Islamic Benchmarking measures, Standards specifies.
5. Islamic Benchmarking
Take JaleesBench and run with it.
First deliverable: JaleesBench v2 and beyond — new scenarios past the current 140, additional pressures and framings, judge-agreement improvements, the results browser, dataset and rubric maintenance, and the cross-tradition family the paper commits to. The construct is faith-general; Islam is instance one.
Explicit responsibility: serving as the evaluation gate for Ansari 4 and Post-training for AI Sovereignty is part of the job, not a favour.
Readiness: this one charters first — volunteers are already asking.
6. Standards for Islamic AI
For something to call itself an Islamic AI assistant, what criteria should it meet?
First deliverable: a published standard — versioned, versionable, and testable wherever possible. Sourcing and citation requirements (does it cite real hadith, and classical sources?), knowledge floors, professed-values alignment, conduct under pressure, disclosure requirements (madhab positioning, “not a mufti” boundaries, when to refer to a scholar), and data and privacy expectations. Possibly tiered — self-attested, benchmarked, reviewed — rather than binary.
Start with: a definition workshop. Is this a standard (measurable criteria) or a guidance document (what builders should consider)? The answer changes the deliverable.
Scope discipline: standards are where volunteer groups most easily become debating societies. Version 0.1 covering assistants only — not all “Islamic AI applications” — is the right first scope.
7. Policy
Shape IASER’s policy work on AI where it touches Muslim communities.
The charter is deliberately open. The co-leads shape it with the founder rather than inheriting a fixed plan. The option space includes: reactive — policy briefs and submissions when regulation touches Muslim communities; proactive — a policy agenda drawn from IASER’s five principles; service — governance toolkits for Muslim organizations adopting AI; or convening — an annual policy convening rather than a standing output.
Charters last, once its co-leads and the founder have shaped the deliverable. It launches with the rest as forming.