Applied AI research, built to withstand scrutiny.
Softforge Labs runs as a structured R&D organization: hypothesis-driven, documented, and reviewed — consistent with the discipline required under France's CIR and JEI innovation frameworks, and with the scientific rigor our behavioral-science and clinical collaborators expect.
Our anchor research program: the cognitive entity
Our central, ongoing research question is how an AI actor should be designed — architecturally and behaviorally — to represent an organization responsibly across six categories of relationship: its own people, its clients, its suppliers, its alliance partners, its customers, and other AI or cognitive systems. This spans open technical problems (identity and permissioning across organizational boundaries, continuity of memory and context, protocol-grounded agent-to-agent coordination) and open behavioral-science problems (trust, explainability, safety boundaries, and dignity in human-support contexts).
Proactive Adaptive Intelligence: our core research thesis
Most AI today is reactive: a person asks, a system answers. Our research asks a different question — can an AI system recognize meaningful change early, build an honest picture of what’s happening, and take the smallest useful action to help a person, an AI agent, or a physical system move toward a better outcome? We call this research domain Proactive Adaptive Intelligence, and it runs on one continuous cycle: Sense, Model, Shape, Measure, Learn.
How can AI recognize meaningful patterns early enough, understand them well enough, and intervene intelligently enough to improve what happens next?
Behavioral & Operational State Intelligence
Understanding what’s actually happening.
We turn raw signals — conversation, timing, telemetry, sensor data — into an honest, evolving picture of state, for a person, an AI agent, or a piece of infrastructure. A single data point rarely means much; a pattern over time usually does.
Adaptive Intervention & Computational Microshaping
The smallest action that helps.
Rather than waiting for a problem to compound and then applying a large fix, we favor small, well-timed interventions that nudge behavior or operations toward a better state — and we measure whether each one actually worked.
Multi-Agent & Multi-Actor Intelligence
Coordinating people, AI, and machines together.
Real environments mix humans, AI agents, and connected systems, each with different authority and expertise. Our research asks who should act, who should defer, and how that coordination should work when the actors aren’t all the same kind of thing.
Outcome Learning & Continuous Optimization
Learning from what actually happened, not what we hoped would happen.
Every intervention is measured against a real outcome, not assumed to work. Results feed back into better decisions next time — for one person, one node, or the whole system.
Trustworthy Autonomy, Authority & Governance
Earning the right to act on its own.
We treat governance as something designed into the architecture, not added on afterward: explicit rules for what a system may decide alone, what needs a human’s sign-off, and an audit trail connecting every action back to what was observed and why.
One research architecture, two very different proving grounds
We test these five pillars in two live, deliberately contrasting environments. Unisia is human-centered: an expert-guided AI companion platform. AmpsPump is infrastructure-centered: a distributed energy platform coordinating charging, storage, and grid interaction across sites. Applying the same research architecture to a human relationship and to a physical energy network isn’t a coincidence — it’s the test of whether the underlying research actually generalizes, rather than being reinvented for every product.
| Unisia | AmpsPump | |
|---|---|---|
| The actor being supported | A person, a team, or an expert-guided relationship | An energy node, a site, or a network of sites |
| What we observe | Conversation, timing, task follow-through, engagement | Utilization, demand, energy conditions, system health |
| What “the right action” looks like | A well-timed prompt, question, or piece of guidance | An adjustment to scheduling, allocation, or routing |
| Who stays accountable | The person, plus an expert who can review and adjust | Operational authority and defined safety limits |
Where this shows up in Unisia
State intelligence models a person’s engagement, hesitation, or follow-through over time. Microshaping means asking for a specific commitment instead of a vague reminder. Multi-actor intelligence coordinates a person, their team or family, an expert, and an AI companion — each with different authority. Outcome learning tracks whether a nudge actually changed behavior days later, not just whether it was well received in the moment. Governance means the person knows adaptive support is active, and an expert can review why the system acted.
Where this shows up in AmpsPump
State intelligence turns live energy telemetry into a read on utilization, congestion, and emerging risk. Microshaping means adjusting charging schedules or power allocation incrementally, before a small inefficiency becomes an outage. Multi-actor intelligence coordinates energy nodes across a site and a region, including how they support each other when conditions change. Outcome learning means results from one site improve the shared model every other site runs on. Governance means the system keeps operating safely and predictably even when it can’t reach central coordination.
Two different domains. One accountable research architecture. That’s the strategic bet behind our R&D program.
This research reaches organizations through Professional Services
When a funded, well-defined opportunity justifies it, this research becomes a deployed engagement — never the other way around.
See Professional Services