Contribution Prerequisites
Target Lifecycle Stage
Incubating - #422
FINOS Member Organization Name
Morgan Stanley
Name
Vexum
Slug
vexum
Business Problem & Solution
▶ Live demo · 📖 Docs · ⌨ Repo · Apache-2.0
The decisions that move money and risk are made in meetings — live, spoken, gone when the call ends; what survives is late, partial notes.
Firms can't point AI at this to fix it: too sensitive to leave their boundary, too risky to let an agent act unsupervised.
So the data → signal → action loop stays slow and lossy.
Vexum closes that loop in real time, inside the firm's boundary.
A bot joins Google Meet / Zoom / Teams natively and captures real-time, speaker-attributed transcripts (Whisper included).
Sandboxed CLI coding agents — the proven coding-agent loop (Codex, Claude Code, etc.) pointed at knowledge instead of code — work the live transcript into the firm's own workspace:
Knowledge as Code — versioned files agents read, write, and commit like a codebase — live during the call and after, which human in the loop.
Safe and scalable — each agent runs in an isolated, ephemeral container (no egress except brokered tools; thousands in parallel).
Deployable in regulated FS — self-hosted, air-gappable, identity-scoped, with a FINOS CALM model in-repo.
Meetings and Agents are separate composable domains, each with its own API — run either alone or together.
Mission Statement
Compound an organization's intelligence by implementing the knowledge-as-code paradigm on safe, scalable infrastructure with first-class access to meetings, while keeping the human in control of the action.
Additional Information (Optional)
Existing materials
- Source: https://github.com/Vexa-ai/vexa-core (Apache-2.0, public,
main) — to be renamed vexum on/before transfer to the FINOS org.
- Docs: https://docs.core.vexa.ai (overview · concepts · architecture · API · deployment · how-to · roadmap · SDKs)
- Live demo: https://core.vexa.ai (the "Vexa Terminal" workbench, on Vexum 0.12)
- Deploy:
git clone … && make all — full stack via Docker Compose on a single Linux host; air-gapped + BYO-inference options documented.
Current state:
-
Proven in production (pre-existing Vexa): the Runtime (meeting bots)
-
Built & proven live — the dispatch core: a unit.v1 dispatch runs in a runtime-spawned isolated container over a bind-mounted workspace, carrying a per-dispatch signed identity token, streaming events to SSE; the in-process path is retired; chat memory is durable. Verified end-to-end on Docker and through the terminal chat.
-
Designed & frozen: the primitives (workspace · agent · container · identity · scheduler), the identity layer (kagenti-aligned), the governance model.
-
On the roadmap: auth-spine + owner-checks, bucket/transcript/user-token encryption, Calendar, Email integraion. Contract-conformance gates + golden fixtures ship in-repo.
Dependencies & licenses: all permissive (Apache-2.0/MIT). STT via faster-whisper + CTranslate2 (MIT); Whisper large-v3-turbo weights (MIT) are downloaded at runtime, not redistributed. No copyleft in the contributed tree.
Adoption & interest
-
Financial-services active use — a Eurosystem central bank: active customer; the headline FS adopter + detailed adoption story (self-host · data sovereignty · GDPR), available private-to-TOC. (Name withheld pending the institution's consent to be referenced publicly.)
-
Member-org interest — Morgan Stanley via Peter Smulovics (TOC Vice Chair), confirmed self-host interest.
-
Broader open-project adoption (proves standalone open-core value + long-term value, non-FS): Infobip (CPaaS unicorn) self-hosts in production K8s and has contributed 4 PRs; Sony · Disney · ILM build on it via the Academy Software Foundation's DNA project (ASWF-DNA participants, not direct customers).
Maintainer Team
| Name |
Affiliation |
Email |
GitHub Username |
| Dmitriy Grankin |
Vexa |
dmitry@vexa.ai |
@DmitryG228 |
Compliance & Requirements Agreement
Administrative Consent & Transfer Authorization
Socialization Acknowledgement
Contribution Prerequisites
Target Lifecycle Stage
Incubating - #422
FINOS Member Organization Name
Morgan Stanley
Name
Vexum
Slug
vexum
Business Problem & Solution
▶ Live demo · 📖 Docs · ⌨ Repo · Apache-2.0
The decisions that move money and risk are made in meetings — live, spoken, gone when the call ends; what survives is late, partial notes.
Firms can't point AI at this to fix it: too sensitive to leave their boundary, too risky to let an agent act unsupervised.
So the data → signal → action loop stays slow and lossy.
Vexum closes that loop in real time, inside the firm's boundary.
A bot joins Google Meet / Zoom / Teams natively and captures real-time, speaker-attributed transcripts (Whisper included).
Sandboxed CLI coding agents — the proven coding-agent loop (Codex, Claude Code, etc.) pointed at knowledge instead of code — work the live transcript into the firm's own workspace:
Knowledge as Code — versioned files agents read, write, and commit like a codebase — live during the call and after, which human in the loop.
Safe and scalable — each agent runs in an isolated, ephemeral container (no egress except brokered tools; thousands in parallel).
Deployable in regulated FS — self-hosted, air-gappable, identity-scoped, with a FINOS CALM model in-repo.
Meetings and Agents are separate composable domains, each with its own API — run either alone or together.
Mission Statement
Compound an organization's intelligence by implementing the knowledge-as-code paradigm on safe, scalable infrastructure with first-class access to meetings, while keeping the human in control of the action.
Additional Information (Optional)
Existing materials
main) — to be renamedvexumon/before transfer to the FINOS org.git clone … && make all— full stack via Docker Compose on a single Linux host; air-gapped + BYO-inference options documented.Current state:
Proven in production (pre-existing Vexa): the Runtime (meeting bots)
Built & proven live — the dispatch core: a
unit.v1dispatch runs in a runtime-spawned isolated container over a bind-mounted workspace, carrying a per-dispatch signed identity token, streaming events to SSE; the in-process path is retired; chat memory is durable. Verified end-to-end on Docker and through the terminal chat.Designed & frozen: the primitives (workspace · agent · container · identity · scheduler), the identity layer (kagenti-aligned), the governance model.
On the roadmap: auth-spine + owner-checks, bucket/transcript/user-token encryption, Calendar, Email integraion. Contract-conformance gates + golden fixtures ship in-repo.
Dependencies & licenses: all permissive (Apache-2.0/MIT). STT via faster-whisper + CTranslate2 (MIT); Whisper
large-v3-turboweights (MIT) are downloaded at runtime, not redistributed. No copyleft in the contributed tree.Adoption & interest
Financial-services active use — a Eurosystem central bank: active customer; the headline FS adopter + detailed adoption story (self-host · data sovereignty · GDPR), available private-to-TOC. (Name withheld pending the institution's consent to be referenced publicly.)
Member-org interest — Morgan Stanley via Peter Smulovics (TOC Vice Chair), confirmed self-host interest.
Broader open-project adoption (proves standalone open-core value + long-term value, non-FS): Infobip (CPaaS unicorn) self-hosts in production K8s and has contributed 4 PRs; Sony · Disney · ILM build on it via the Academy Software Foundation's DNA project (ASWF-DNA participants, not direct customers).
Maintainer Team
Compliance & Requirements Agreement
Administrative Consent & Transfer Authorization
Socialization Acknowledgement