Implementation roadmap
Version française : ../fr/roadmap.md
Twelve phases deliver the 8 bricks (A identity and governance, B LLM management, C agents and execution, D tools and integrations, E data, RAG and memory, F multimodal, G quality, observability and audit, H FinOps, infrastructure and operations). Brick pages: docs/en/bricks/.
| Phase | Priority | Scope | Bricks |
|---|---|---|---|
| 1 | critical | Contracts, architecture, identity, tenants, policies | A-01..A-14, G-11, H-06, H-07, H-10, H-16 |
| 2 | critical | LiteLLM, vLLM, providers, model registry, quotas | B-01..B-07, B-09..B-11, B-14, H-01, H-02, H-03 |
| 3 | critical | Agent runtime, Python SDK, tool gateway, MCP | C-01..C-04, C-06..C-08, C-11..C-14, D-01..D-04, D-14 |
| 4 | critical | DLP, encryption, tokenization, RAG ACL | E-01..E-04, E-06..E-08, E-11..E-13 |
| 5 | critical | OpenTelemetry, Opik, audit, ClickHouse | G-01..G-08, G-11, G-12, G-14 |
| 6 | high | Quality gates, Promptfoo, Ragas, red teaming | B-12, G-09, G-10, G-13, H-15 |
| 7 | high | Skills, plugins, connectors, channels | D-05..D-13, H-12, H-13 |
| 8 | high | Context, memory, sub-agents, durable execution | B-08, C-05, C-09, C-10, C-15, E-09, E-10 |
| 9 | high | Vision, images, documents, STT, TTS, voice live | F-01, F-02, F-05..F-12, E-05 |
| 10 | high | Video, multimodal generation, GPU scheduling | F-03, F-04, B-13, H-04, H-05 |
| 11 | critical before production | HA, DRP, load tests, security, compliance | H-06..H-16, A-13, A-14, G-11, G-12 |
| 12 | continuous | Advanced FinOps, optimization, quality improvement | H-01..H-04, B-09..B-11, G-04, G-07, G-09 |
Rule: phase 11 requirements apply from phase 1
Phase 11 is the production gate, not the moment security and resilience start. From phase 1 every deliverable already respects:
- tenant isolation (RLS), least privilege roles and scopes, secrets only in OpenBao, no secret in logs;
- hash chained audit of every mutation and policy decision;
- stateless services with probes, PDBs, resource limits and at least 2 replicas in the Helm chart;
- idempotent migrations and APIs (Idempotency-Key), backup ready data layout;
- images built in CI, scanned (Trivy), with SBOM (Syft), signed (Cosign), pulled from Harbor by digest;
- OpenTelemetry instrumentation and Usage Events for anything that costs;
- bilingual (en/fr) errors and UI copy, no em dash or en dash.
Phase 11 then proves these properties at scale (load, chaos, DR drill, pentest, compliance evidence).
Phase 1: Contracts, architecture, identity, tenants, policies (critical)
Deliverables. 15 JSON Schema contracts v1, Control Plane API v1, Keycloak OIDC, API keys, SPIFFE identities, tenants/projects/environments with PostgreSQL RLS, policy engine (OPA/OpenFGA with builtin fallback), approvals, hash chained audit, secrets in OpenBao, Helm chart skeleton.
Bricks. A-01..A-14, G-11, H-06, H-07, H-10, H-16
Exit criteria. Every contract validated in CI; cross tenant read impossible (RLS test); every mutation audited and GET /v1/audit/verify valid; policy decisions recorded with decision id.
Phase 2: LiteLLM, vLLM, providers, model registry, quotas (critical)
Deliverables. Provider registry (16+ providers), model registry and statuses, Capability Resolver, Model Router with fallbacks, LiteLLM config generation, vLLM deployments, gateway endpoints, estimate, quota service with Redis reservations, price cards and Usage Events.
Bricks. B-01..B-07, B-09..B-11, B-14, H-01, H-02, H-03
Exit criteria. All model calls go through the gateway endpoints (/v1/gateway/chat/completions and siblings); kill switch on a provider takes effect in under 30 s; parallel calls never exceed a quota; each call produces one Usage Event with estimated and measured cost.
Phase 3: Agent runtime, Python SDK, tool gateway, MCP (critical)
Deliverables. Agent registry and runs (Execution Contract), steps, checkpoints, heartbeats, sandbox executions, Tool Gateway with scopes/policy/DLP/approval, MCP gateway, Python SDK and CLI first release.
Bricks. C-01..C-04, C-06..C-08, C-11..C-14, D-01..D-04, D-14
Exit criteria. A reference agent runs end to end from the SDK with tools and MCP, can be suspended and killed, and every tool call is policy checked and audited.
Phase 4: DLP, encryption, tokenization, RAG ACL (critical)
Deliverables. Classification, text DLP (Presidio + French recognizers), tokenization and detokenization under policy, envelope encryption with OpenBao, knowledge bases with ACL aware retrieval (Qdrant), retention and erasure, lineage.
Bricks. E-01..E-04, E-06..E-08, E-11..E-13
Exit criteria. No PII reaches an external provider when policy forbids it; a user never retrieves a chunk he cannot read at the source; erasure request completed and evidenced.
Phase 5: OpenTelemetry, Opik, audit, ClickHouse (critical)
Deliverables. OTel tracing and collector pipeline, Opik AI tracing, ClickHouse analytics, Prometheus, OpenSearch (logs), Jaeger, Alertmanager, WORM audit export, audit correlation, Wazuh and SIEM forwarding.
Bricks. G-01..G-08, G-11, G-12, G-14
Exit criteria. One trace per request across sync and async paths; audit export stored with Object Lock; dashboards for cost, latency and errors per tenant; alerts routed with runbooks.
Phase 6: Quality gates, Promptfoo, Ragas, red teaming (high)
Deliverables. Evaluation suites and runs (Ragas, Promptfoo), red team runs (garak), quality orchestrator with blocking gates in CI and before prod promotion.
Bricks. B-12, G-09, G-10, G-13, H-15
Exit criteria. No agent or prompt version reaches prod without fresh evaluation and red team evidence; waivers need approval.
Phase 7: Skills, plugins, connectors, channels (high)
Deliverables. Signed skills and Skill Loader, plugins lifecycle, connectors with Credential Broker and webhooks, channels (web, Teams, Slack, telephony).
Bricks. D-05..D-13, H-12, H-13
Exit criteria. Unsigned skills refused; connector credentials never readable; inbound channel messages mapped to runs with identity.
Phase 8: Context, memory, sub-agents, durable execution (high)
Deliverables. Context Manager budgets and compression, memory contract, sub-agents with depth limits, A2A messages, durable runs (PostgreSQL checkpoints, AgentLoop.resume) and schedules run by python -m eai.worker, with an optional Temporal adapter.
Bricks. B-08, C-05, C-09, C-10, C-15, E-09, E-10
Exit criteria. Long runs survive a worker restart from checkpoints; context_overflow never reaches a provider; sub-agent limits enforced.
Phase 9: Vision, images, documents, STT, TTS, voice live (high)
Deliverables. Media service, vision, image generation and editing, STT (faster-whisper), diarization (pyannote), TTS (Piper), VAD (Silero), voice live (LiveKit), audio streaming, multimodal DLP.
Bricks. F-01, F-02, F-05..F-12, E-05
Exit criteria. Voice live p95 latency under 800 ms; all media scanned, classified and DLP checked; Usage Events for images, audio seconds, characters and sessions.
Phase 10: Video, multimodal generation, GPU scheduling (high)
Deliverables. Video processing (FFmpeg/PyAV), video generation GPU workers, GPU pools with KEDA, DCGM based GPU accounting and allocation policies.
Bricks. F-03, F-04, B-13, H-04, H-05
Exit criteria. GPU pools scale to zero when idle; every local GPU second is allocated to a tenant; video jobs resume after preemption.
Phase 11: HA, DRP, load tests, security, compliance (critical before production)
Deliverables. High availability of every stateful component, backup and DR (Velero, PITR) with drills, load and soak tests, penetration test, image scanning, SBOM, signing, admission policies, GitOps, IaC, compliance evidence (AI Act, GDPR, ISO 27001 mapping).
Bricks. H-06..H-16, A-13, A-14, G-11, G-12
Exit criteria. RPO and RTO met in a drill; load test at target throughput with SLOs green; no critical vulnerability; signed images only; evidence package exported.
Phase 12: Advanced FinOps, optimization, quality improvement (continuous)
Deliverables. Reconciliation with invoices, variance alerts, chargeback, routing to cheaper models under quality constraints, prompt caching optimization, quality trend reviews.
Bricks. H-01..H-04, B-09..B-11, G-04, G-07, G-09
Exit criteria. Measured vs billed variance below threshold every month; cost per successful task trending down at constant quality.
Already implemented in this repository
State checked in the working tree on 2026-10-29 (served by GET /v1/roadmap and GET /v1/catalog/bricks). Phase 11 requirements apply from phase 1: every item below already ships with RLS, chained audit, en/fr i18n, probes and resource limits in the chart.
| Phase | Implemented |
|---|---|
| 1 | contracts/v1 (15 contracts + common), contracts/validate.py and contracts/tests; domains system, contracts, identity, tenancy, authz, policies, approvals, secrets, governance, risk, incidents; migrations sql/postgres/V001 to V011; portal screens admin/*, security/*, audit/*; deploy/keycloak, deploy/policies (OPA with tests, OpenFGA, Kyverno), deploy/openbao, deploy/spire, deploy/cert-manager, Helm chart |
| 2 | Domains llm (providers, models, routes, resolver, context, gateway, vLLM, LiteLLM) and finops (prices, usage, quotas, GPU, reconciliation, alerts); migrations V020 to V035; screens llm/*, finops/*; deploy/litellm, chart vllm and litellm values |
| 3 | Domains agents, tools, mcp; Navin runtime of the Python SDK (enterprise_ai.loop.AgentLoop, enterprise_ai.runtime.AgentRunner, enterprise_ai.subagents.SubagentManager, enterprise_ai.loop.plan.RunPlan, enterprise_ai.runtime.bus.MessageBus) and the enterprise-ai CLI; worker python -m eai.worker; migrations V040 to V042; screens agents/*, capabilities/tools, capabilities/mcp; sandbox (gVisor and Kata classes) |
| 4 | Domains privacy (classification, DLP, tokenization, encryption), knowledge, retention, lineage; migrations V060 to V063; screens data/*; Presidio, Qdrant, OpenBao transit adapters |
| 5 | Domain observability and audit (export, verify, correlations); sql/clickhouse; deploy/observability (OTel Collector, Prometheus, OpenSearch, Jaeger, Opik, Alertmanager), deploy/wazuh; migration V065 |
| 6 | Domain evaluation (suites, runs, gates, red team) with Ragas, Promptfoo, garak adapters; deploy/ci; screens quality/* |
| 7 | Domains skills, plugins, connectors, channels; screens capabilities/* |
| 8 | Domains memory, prompts, sessions, A2A, schedules and triggers (tick and outbox in eai.worker); migration V062 |
| 9 and 10 | Domains multimodal and realtime (vision, image, video, STT, diarization, TTS, LiveKit); migration V064; screens multimodal/*; deploy/keda, deploy/docker/worker.Dockerfile |
| 11 | deploy/velero, deploy/terraform, deploy/argocd, deploy/gitops, deploy/cilium, deploy/loadtest (k6), values-prod.yaml, docs/operations (HA/DR, hardening, runbooks, phase 11 readiness) |
| 12 | Reference price cards (V035), reconciliation and variance alerts |
SDKs: Python (sdks/python), TypeScript (sdks/typescript/packages/sdk and realtime), Java (sdks/java, Spring module included), Go (sdks/go, packages realtime, worker, gatewaymw, usage); they cover the paths of docs/api/control-plane-v1.md.
Remaining before production
- Integration tests against real Keycloak, OpenBao, OpenFGA, Qdrant, NATS, Temporal, LiveKit, LiteLLM and vLLM deployments (tests currently use a real PostgreSQL and the in-process fallbacks of the adapters).
- Execution of the
deploy/loadtestload tests at target throughput, with measured SLOs. - External penetration test.
- Verification of the reference price cards (
V035__finops_seed_reference_prices.sql) against the official provider price lists. - DR drill (Velero restore and PostgreSQL PITR) with measured RPO and RTO.
Brick pages F, G, H
See docs/en/bricks/F-*.md, G-*.md, H-*.md. Cost rules (estimated, measured, billed) are detailed in H-02 cost ledger and H-03 pricing engine.