Osmoze, by Heavenize

AI, without changing your architecture.

We do not build experimental AI agents. We provide a goal-oriented orchestration engine, proven over 15 years in capital markets, that LLMs plug into to bring a natural interface and reasoning capabilities.

A ready-to-use AI cockpit for common Front Office use cases, extensible via the framework for your specific use cases: an agent that calls your business building blocks as tools, under a single audit trail.

Traceable, explainable, compatible with your compliance requirements.

A transparency note on AI. The Intentional Engine, the orchestration engine at the heart of Osmoze, has been in production at our clients since 2012. The AI layer that plugs into it, described below, has just been delivered in the product: proven architecture, AI capabilities in early commercial rollout.

We are identifying one or two reference clients for this AI layer: you co-define the use case, we certify it together. It's also another angle worth discussing.

Why AI projects in
asset management stall

For three years, asset management firms have launched dozens of AI Proofs of Concept. How many in production one year later? Very few.

The projects that fail share the same causes: requirements written far away from the portfolio manager's workstation, AI portals deployed next to Excel rather than inside it, generic models that understand neither the internal conventions nor the specificities of a portfolio management team. The result: a nice demo, and three months later no one uses it any more.

For AI to take hold in the Front Office, it must: live in the tool portfolio managers already use (Excel), speak the team's language (conventions, indicators, workflows), rely on real portfolio data, and remain explainable, auditable and compliant. No magic, no black box.

This is precisely the observation that guided the design of Osmoze's AI: make it live where it has access to data, calculations and the portfolio manager's workflow, inside the same architecture.

AI inside the same architecture

An AI layer that runs inside the Osmoze engine, not alongside it.

Osmoze is not just a collection of pre-packaged agents. It is a ready-to-use AI cockpit for common Front Office use cases, and a framework (a specialised development toolkit) for building your own agents when your need goes beyond the standard scope. AI does not live in a separate portal: it runs in the same architecture as the business calculations, accesses the data hub directly, and inherits the same audit trail.

Direct consequences: an agent can call business building blocks (pricing, compliance, stress test) as tools, see data in context, and explain its actions, sources, calculations and justifications down to the source data. And a new agent can be built, calibrated and deployed in a few days on a focused use case, when the industry typically measures this in weeks or even quarters.

Any agent, quickly

A new need in a team? The corresponding agent can be specified, calibrated and tested in a few days. No heavy redesign, no endless specification cycle: the framework absorbs iterations without breakage.

Natively connected to Osmoze

The agent accesses the full Osmoze hub (inventories, positions, transactions, history, news, internal documents) and calls the business calculation libraries directly (pricing, stress, performance contribution, optimisation).

Model-agnostic architecture

Compatible with all major LLM providers (hosted or local), with the capability to fine-tune open-source models on your data without it ever leaving your infrastructure.

Traceable, explainable, compatible with your compliance

Our signature. Every result produced by an agent can be traced back to its data sources and to the calculation chain that produced it, directly explained to the user. This is both a condition for adoption by portfolio managers and a response to regulatory requirements on control and auditability of AI models in asset management.

A few concrete examples

Five agents built on the Osmoze framework to illustrate the spectrum: capabilities available in the latest release, in early commercial rollout

These agents are not a closed catalogue. They are illustrations of what can be built quickly on the framework. As soon as a new idea appears in your team, it becomes a possible agent.

Morning Briefing

A morning brief personalised per portfolio. News pre-filtering on held securities, weak-signal detection, macro summary. Reading time: about twenty minutes instead of an hour of scattered monitoring.

Customiser

Generates indicators, analysis axes or compliance rules on the fly as soon as a new need appears. The tool grows at the portfolio manager's pace, without an IT ticket.

Investment Comments Generator

Writes a fund's portfolio management comments from internal data (contribution, transactions, inventories) and external data (news, sectors). Respects the style and continuity of previous comments.

Portfolio News Impact Evaluator

Assesses the impact of the latest market information on a specific portfolio. Combines the inventory, the exposures computed by Osmoze, and an AI search engine for external sources.

Enterprise Expert

An LLM fine-tuned on your firm's knowledge and practices: RFPs, analyses, sector summaries. No sensitive document is sent outside: the model lives inside your infrastructure.

Your next agent

A specific use case in your team? A recurring report that takes time? An analysis you no longer do for lack of bandwidth? That is exactly the kind of need the framework turns into an agent in a few days.

Two families of AI
in a single cockpit

Proprietary Deep Learning

For structured and quantitative data.

We have in-house deep learning models, trained on twenty years of market data. They apply to:

  • Pattern and weak-signal detection
  • Prediction of parameters or curve evolution
  • Constrained optimisation and temporal planning

This know-how is at the heart of our Autopilot roadmap.

Large Language Models (LLM)

For unstructured data: text, documents, multimodal.

Open multi-provider architecture, with fine-tuning capability on your internal data:

  • Summarising reports, prospectuses, news
  • Generation of business documents
  • Qualitative analysis and contextual reasoning

Sensitive models stay local; your data only leaves if you choose to allow it.

In Osmoze, both families coexist, and a single agent often combines the two, on the same data foundation and with the same traceability.

For your RFI / IT department Regulatory and technical detail on compliance and AI audit

Referenced regulatory compliance

What "compatible with your compliance requirements" concretely means.

Osmoze is designed to fit within the regulatory frameworks of European asset management firms and insurance companies. Compatibility with the requirements below is documented and defendable in RFI / tender / IT due diligence:

Solvency II

Prudential calculations integrated (market SCR, cash-flow projection, yield curves). Compatible with the QRT format and ACPR reporting flows.

DORA (Digital Operational Resilience Act)

Incident traceability, IT third-party risk management, critical providers registry. Architecture compatible with a full DORA audit.

EU AI Act (articles 13-14)

Technical documentation of AI systems, decision traceability, human oversight mechanisms (the portfolio manager validates, the agent proposes). Inference logs retained.

AMF / RG AMF

Compatibility with the Front Office obligations of French asset management firms: ex-ante / ex-post compliance, decision traceability, archival of decision contexts (safety memory).

What the AI audit trail contains

At the level of detail expected in an RFI by your CISO.

By architectural design, for every interaction with an Osmoze AI agent, the audit trail is designed to retain:

  • Model identity: name, version, provider, hosting (public cloud / private / on-premise)
  • User prompt: the question asked by the portfolio manager, in clear text
  • Successive tool calls: which tools the agent invoked, in what order, with what parameters
  • Data sources accessed: portfolios, indicators, business calculations, reference data, each timestamped
  • Intermediate responses: what each tool returned to the agent
  • Final output: the response returned to the portfolio manager, with its citations
  • Execution context: user, rights, scope, end-to-end timestamps
  • Data hashes: cryptographic fingerprint of positions and calculations at the time of inference, for later deterministic replay

Reproducibility: from this recording, the agent's decision can be replayed identically months or years later, with the exact model version and the exact state of the data, which is rarely true in standard AI implementations on the market.

Advanced R&D

Portfolio Autopilot

Our next step, in advanced R&D. A proprietary Deep Learning layer, in the same architecture as the Intentional Engine, which simultaneously integrates target allocation, available cash, compliance constraints, risk profile and market constraints. The portfolio manager stays the decision-maker: the agent proposes, the portfolio manager validates. First reference clients in 2026.

A concrete use case in mind?

Every Osmoze agent is built from a real client need. Share yours with us, even if it is still vague or in the exploration phase.

Book a meeting