Agentic Signal Classification Engine on Intel Xeon 6
Three-tier agent cascade that classifies enterprise signals into governed, verified action. Deterministic nanoagents compress on CPU. LLM reasoning when you need it. Agents earn their tier through empirical validation.
"The first job of enterprise AI is not generation. It is classifying reality well enough to know what should happen next."
Signals → Decide → Act → Verify → Learn
Historical / Live / Synthetic Sources
│
▼
Evidence Normalizer
│
├────────────────────┐
▼ ▼
Baseline Compiler Runtime Signal Flow
│ │
▼ ▼
Baseline Profiles Nanoagent Classification (7 agents, Intel Xeon 6)
│ │
└─────────┬──────────┘
▼
Microagent Inference (5 agents, CPU / LLM)
│
▼
Macroagent Reasoning (5 agents, CPU / Gaudi)
│
▼
Decide → Act → Verify → Learn
│
▼
Dashboard + API + Bootstrap Lab
Three-tier classification cascade:
- Nanoagents (7) — Deterministic, no LLM, Intel Xeon 6. Baseline distance, metric drift, log patterns, document heuristics, image/audio metadata, evidence gating. Zero inference cost.
- Microagents (5) — Rule-backed classifiers on CPU. Text, document, image defect, audio anomaly, embedding clustering. Optional LLM via Granite 3.2 8B. Extension points for Intel OpenVINO/ONNX.
- Macroagents (5) — Higher-level reasoning. Incident timeline, root cause hypothesis, action planning, verification planning, learning proposals. Template-based on CPU, or LLM-backed via Gaudi/Xeon.
Image/audio microagents are fixture-backed by default for a dependency-free demo. Set DEEPFIELD_MEDIA_BACKEND=onnx with DEEPFIELD_IMAGE_ONNX_MODEL or DEEPFIELD_AUDIO_ONNX_MODEL to enable optional CPU media adapters.
Agent Promotion Pipeline:
- Agents start as draft and earn their tier through empirical validation
- Draft → Candidate (50 samples, 60% accuracy) → Nano (200 samples, 75%) → Micro (500 samples, human reviewed) → Macro (1000 samples, cross-modal agreement)
- Red/yellow/green rubric matrix tracks every agent's maturity
- Only promoted (green) agents run in the active pipeline
Agent Loop:
- Actions — Propose/approve/execute safe actions (notify, observe, ticket). Non-destructive by design. Human approval gates.
- Verification — Compare post-action observations to expected outcomes.
- Learning — Propose threshold/rule updates. Never applied silently — always reviewed.
# Backend
pip install -e ".[dev]"
pytest app/tests/ -v # 217 backend tests
uvicorn app.main:app --reload
# Frontend
cd frontend
npm install --legacy-peer-deps
npm run dev # http://localhost:3000 (proxies to :8000)
# Container
podman run -p 8000:8000 quay.io/redhat-gpte/deepfield-multimodal:latest
# CLI demo (no server needed)
python3 -m app.demo
# Measured proof report
python3 -m app.benchmark --profile enterprise-signal-volume --iterations 5 --include-project-tests --out benchmark-results/latest.json
# Health check
curl http://localhost:8000/health| Section | Duration | What happens |
|---|---|---|
| Presentation | ~5 min | 7 click-through slides — business case, measured CPU compression, three tiers |
| Walkthrough | ~10 min | 6 manual acts — ingest, baseline, nano/micro/macro cascade, act, learn |
| Scale Run | ~5 min | 13 auto steps — 10→50 lines, stress test, recovery, the claim |
| Bootstrap Lab | ~20 min | Pick scenario → analyze → validate → rubric matrix → promote agents |
Four synthetic scenarios for self-paced labs:
| Scenario | Domain | Signals | Profile |
|---|---|---|---|
| OpenShift Cluster Health | IT Ops | 156 (pods, events, nodes) | openshift-monitoring |
| Factory Floor Monitoring | Manufacturing | 6 (vibration, temp, logs, image, audio) | — |
| Telecom Network Operations | Telecom | 150 (signal strength, events, logs) | — (frontier model optional) |
| AAP Job Failures | IT Ops | 100 (jobs, workflows) | aap-job-health |
Two analysis paths:
- Quick Start — pre-built profile, instant, no LLM needed
- Deep Analyze — frontier-model semantic analysis when configured, generates domain-specific rules
| Tier | Model | Hardware | When |
|---|---|---|---|
| Nano (runtime) | None — deterministic rules | Intel Xeon 6 CPU | Every signal, always |
| Micro (runtime) | Granite 3.2 8B (optional) | Intel Xeon 6 CPU | Escalated evidence only |
| Macro (runtime) | Granite 3.2 8B (optional) | Intel Xeon 6 / Gaudi 3 | Cross-modal correlation |
| Bootstrap (one-time) | Qwen 3 235B | Intel Gaudi / MaaS | Initial data analysis |
98% of signals classified on CPU before anything expensive runs. Verified by benchmark CLI at 100% in rule-backed mode.
| Route | Description |
|---|---|
GET /health |
Health check |
| Demo | |
POST /api/v1/demo/start |
Start auto-run demo |
GET /api/v1/demo/state |
Poll demo state (SSE at /api/v1/stream) |
GET /api/v1/demo/infrastructure |
Runtime + agent inventory |
| Benchmark | |
GET /api/v1/benchmark/latest |
Latest measured CPU-compression report |
POST /api/v1/benchmark/run |
Run benchmark profile and optionally save report; pass include_project_tests: true for backend/frontend validation |
| Bootstrap | |
GET /api/v1/bootstrap/scenarios |
List lab scenarios |
POST /api/v1/bootstrap/scenarios/{id}/load |
Load scenario data |
GET /api/v1/bootstrap/profiles |
List pre-built profiles |
POST /api/v1/bootstrap/profiles/{id}/apply |
Apply profile (no LLM) |
POST /api/v1/bootstrap/connect |
Connect live data source |
POST /api/v1/bootstrap/analyze |
Semantic analysis (Qwen/Sonnet) |
POST /api/v1/bootstrap/validate |
Run validation round |
GET /api/v1/bootstrap/rubric |
Agent maturity rubric matrix |
POST /api/v1/bootstrap/promote/{id} |
Promote agent (human review) |
| Classification | |
POST /api/v1/classification/run |
Run classification cascade |
POST /api/v1/demo/classify/nano |
Nano tier only |
POST /api/v1/demo/classify/micro |
Micro tier only |
POST /api/v1/demo/classify/macro |
Macro tier only |
# OpenShift with OAuth proxy
oc apply -f deploy/deployment.yaml
# Verify (13 checks)
bash deploy/verify.shContainer: quay.io/redhat-gpte/deepfield-multimodal:latest
Requires: cluster-reader + cluster-monitoring-view ClusterRoles on ServiceAccount.
| Check | Grade |
|---|---|
| NovaScan | Partner / Self-Serve / $0 per session |
| DarkScope | A — 0 findings, score 0 |
| Brand Audit | A — 155/170, Intel + Red Hat aligned |
| Preflight | READY |
CDD → TDD → BDD → EDD
- CDD — Contracts defined as Pydantic models and function signatures
- TDD — Tests written RED first, then implemented to GREEN
- BDD — Given/When/Then scenario tests for end-to-end flows
- EDD — Rubric scoring (healthy/warning/failing) across quality dimensions
217 backend tests plus frontend component tests and production build validation. 9 EDD rubric dimensions. All green.
deepfield-multimodal/
├── app/
│ ├── domain/models.py # 12 Pydantic models (CDD contracts)
│ ├── multimodal/ # Normalizer, feature extractors, storage, scale generator
│ ├── baseline/ # Compiler, profiles, sources
│ ├── classification/ # Engine, taxonomy, cascade, registry
│ ├── nanoagents/ # 7 deterministic agents + pipeline
│ ├── microagents/ # 5 rule-backed + 1 configurable (LLM)
│ ├── macroagents/ # 5 reasoning agents
│ ├── agent_loop/ # Actions, verification, learning, orchestrator
│ ├── bootstrap/ # Semantic classifier, promotion, constraints, scenarios
│ ├── connectors/ # File, Prometheus, Kubernetes
│ ├── inference/ # LiteLLM client (runtime + bootstrap)
│ ├── benchmark.py # Measured CPU-compression benchmark CLI/API engine
│ ├── analysis/evaluator.py # EDD rubric scoring engine
│ ├── api/ # 7 FastAPI routers + SSE streaming
│ └── tests/ # 217 backend tests (CDD/TDD/BDD/EDD)
├── frontend/ # React 19, motion/react, inline styles
├── fixtures/ # Factory scenario + 4 lab scenarios
├── config/ # YAML configs (taxonomies, profiles, promotion thresholds)
├── deploy/ # OpenShift manifests + verify.sh
├── agnosticv/ # RHDP catalog config
├── docs/ # Antora documentation and presenter/lab guides
└── migrations/ # PostgreSQL schema (optional)
Red Hat OpenShift · Intel Xeon 6 · Intel Gaudi 3 · Intel TDX