Feneri is the multimodal foundation model for human signal: real-time emotion, attention, and intent from video, voice, and text. The substrate beneath every agent, robot, game, and clinical product that needs to read the room.
Agents do not notice frustration. NPCs do not react to a player's mood. Robots do not sense when a human is uncomfortable. Voice assistants cannot tell distress from sarcasm. Feneri is the layer that fixes this. One model, one API, every builder.
We started Feneri because every AI breakthrough of the last five years has made machines smarter, not more attentive. The next class of systems (agents, robots, clinical companions) will fail the same way today's models do, until they can sense the person on the other side of the glass. That's the model we're building.
The Feneri MMLM fuses facial micro-expressions, vocal biomarkers, and linguistic sentiment into a single multimodal signal, richer than any one input, deployed sub-second, bilingual Arabic-English from day one. Trained on 4.2 billion frames, owned end-to-end. No wrappers, no third-party models.
Three input streams enter a 12-head multimodal fusion core. A single state object falls out the other end: emotion, attention, intent, every 33 milliseconds, every modality, every region.
However your team builds (direct API calls, native SDKs, MCP for the agent stack, or a no-code session in Studio), the same engine answers, and the same signal comes back.
One /v1/read endpoint. Streaming or one-shot. WebSocket, gRPC, and REST. The platform's primitive.
Native bindings for the surface you build on. feneri-agents, -ros, -game, -clinical, -ops, -edge.
Plug Feneri into any MCP-compatible host. Claude Desktop, Cursor, Cody, custom agents. One command, six tools, instant emotional context.
Upload a recording or run a live session. Get a full affective report, mental-health indicators, and exportable analysis. Zero code.
Native SDKs across every surface you ship on. And when something matters, Feneri calls you back: outbound webhooks for state transitions, not endless polling.
For LLM agents and copilots. TypeScript & Python. Streaming context for any orchestrator: LangChain, LlamaIndex, custom.
ai & agentsFor embodied systems. ROS 2 nodes, edge-quantized models, on-device inference for humanoids and service robotics.
roboticsPlugins for Unity and Unreal. C# & Blueprint bindings. Frame-aligned emotion for NPC behavior trees and adaptive difficulty.
gamingHIPAA-ready bindings for telehealth platforms. Session monitoring, clinician dashboards, triage alerts.
telehealthCCaaS- and CRM-ready. Live agent assist, supervisor alerts, sentiment-driven routing. WebRTC & SIP-bridge.
customer opsSwift, Kotlin, WebAssembly. System-level emotion for OS vendors and device makers. On-device, zero-network mode.
tech & osSubscribe to state-transition events and Feneri posts to your endpoint the moment something matters: stress crosses a threshold, attention drops, distress baseline shifts. Build alerting, escalation, or context-injection without polling.
Six product surfaces. One emotional substrate. Embed the engine in an agent's planner, a robot's behavior tree, a game's session loop, or call the API once per session and route the state wherever you need it.
Give LLM agents real-time affective context. Adjust pacing, detail, and tone the moment a user disengages, or escalate to a human before they churn.
Humanoids and service robots that respond to people, not commands. Edge-ready inference for ROS 2.
Dynamic difficulty, emotion-aware dialogue, frustration detection. SDKs for Unity and Unreal.
HIPAA-ready affective biomarkers for telepsychiatry, CBT outcome tracking, and early-warning intervention.
Detect frustration before churn signals appear in your CRM. Trigger live coaching for the agents who need it.
Wearables, in-car, edge sensors. Ambient affective intelligence as a system service, on-device.
All raw video and audio is permanently deleted within sixty seconds of inference. Only anonymized state metadata persists. We never own, and can never leak, your users' faces or voices.
The /v1/read endpoint accepts video, audio, or text, alone or fused, and returns a structured state object you can route into anything: a game loop, a clinical workflow, a robot's planner, an agent's context.
Fener · فنار, a lighthouse.
For two centuries they kept ships from running aground in fog they could not see through.
Feneri does the same for machines: a steady beacon of human signal, so the next generation of systems can finally tell the difference between someone who is fine and someone who is drowning.
Engine access is currently invitation-only. Tell us what you're building. We read every application within 48 hours.