Skip to content

Latest commit

 

History

History
45 lines (38 loc) · 4.62 KB

File metadata and controls

45 lines (38 loc) · 4.62 KB

Off Grid AI — the platform

Become an intelligent enterprise, without compromising. On-prem · local-first · source-available · data never leaves the box. This is the canonical platform model — the spine of the public source. Visual map: platform-map.html.

The thesis (one sentence)

One governed pipeline is the unit of everything — it governs model access AND data movement under one policy/guardrail/redaction/eval contract; it runs on a smart gateway (routing + edge + its own telemetry); every run and every sync emits into one shared spine (observability · audit · lineage · provenance · cost · compliance), is replayable end-to-end, and an org view rolls the whole estate up.

The layers

  • Org view — roll-up across every pipeline/gateway/consumer: spend & chargeback, all traces/audit, every data flow + lineage graph, drift & data-quality, compliance coverage, fleet health, tenants. Drill into any pipeline's slice.
  • Consumers (bind a pipeline) — Apps · Agents · Chat/Projects · BI (Superset/PowerBI) · external 3rd-parties via a provisioned key. Built by non-technical dept staff in plain language.
  • Pipeline — THE governed unit, owns everything. Does model-access (grounded answers + citations) AND data-movement (connector→warehouse; batch/incremental/CDC). Same peripheries govern both: data-allowlist (hard ceiling), policy (ABAC/OPA), guardrails, PII redaction on the sync path, model evals+golden+drift, data-quality evals (Great Expectations), routing/egress leash, schedule/CDC trigger, provisioned API key, versioning, replay. Per-pipeline lenses = its slice of the spine.
  • Gateway — the smart network+model edge: smart routing (model pick/fallback/load-balance/egress leash) + edge functions (Caddy: WAF, rate-limit, cache, TLS, per-tenant URL) + its OWN observability/logs.
  • Substrate — two engines: the model engine (llama.cpp fleet + cloud, OpenAI-compat — LIVE) and the data engine (Airbyte + Redpanda + ClickHouse + dbt + Great Expectations — LIVE on S2, 2026-07-08; the AWS-data-stack parity map is DATA_PLANE_PARITY.md).
  • The spine (run OR sync = the join key): Observability (Langfuse) · Audit/SIEM (OpenSearch) · Lineage (Marquez) · Provenance + Citations · FinOps/Cost · Regulatory (ISO 42001 / NIST / EU AI Act).
  • Cross-cutting: Identity (Keycloak) · Secrets (OpenBao) · Multi-tenancy · Flags (Unleash) · Fleet/MDM (FleetDM) · Storage (SeaweedFS) · Knowledge/Brain (Qdrant/LanceDB) · Backups/DR.

The OSS stack (permissive licences — moat is ownership + air-gap, not secrecy)

Capability Engine Status
Model gateway llama.cpp fleet + OpenAI-compat live
Warehouse (Snowflake/Databricks-equiv) ClickHouse live on S2
Connectors + ELT (Glue-equiv) Airbyte (300+) live on S2
CDC (DMS-equiv) Debezium (via Airbyte) live on S2
Streaming / MQ Redpanda (Kafka API) live on S2
Transforms dbt job/CLI
Data-quality evals Great Expectations live on S2
PII redaction Presidio live
Model evals / drift ragas · deepeval · Evidently live
Observability Langfuse live
Audit / SIEM OpenSearch live
Lineage Marquez (OpenLineage) live
Policy OPA (ABAC) live
Secrets · Identity · Durable runs OpenBao · Keycloak · Temporal live
Flags · MDM · BI · Object store Unleash · FleetDM · Superset · SeaweedFS live

Experience principle (the product bet)

The platform must feel beautiful and effortless — creating and deploying a governed use-case is the core loop, and it should be a joy, not a config chore. Two pillars:

  1. A LOT of prebuilt pipelines — a rich, curated library of ready-to-run governed use-cases (BFSI: KYC, loan underwriting, fraud, FNOL, reimbursement, cross-sell; ETL: connector→warehouse syncs; plus general). Clone → tweak in plain language → deploy. The library is the fastest path to value and the showcase of the platform.
  2. Plain-language authoring — describe the use-case in English; the platform wires the pipeline (gateway binding, data allowlist, guardrails, steps, schedule). Non-technical dept staff are the acceptance bar.

Fleet

10 machines: S1 control-plane · S2 data-plane (Airbyte/ClickHouse/Redpanda/GE — LIVE under OrbStack, fronted on S1 edge loopbacks 8941–8944) · g1–g8 model nodes.

Positioning

"A private AI, everywhere" + the governed data platform under it. Fair-code: free for up to 25 users; larger deployments require a commercial license.