— Technical Brief

Built for the enterprise. Deployed in minutes.

One container. Your VPC. Your data. Your governance. ShareContext sits alongside the warehouses, databases, and identity providers you already have — and starts producing value before the kickoff meeting ends.

1 container
Single deployable
0 bytes
Data egress
<10 min
Time to first value
model agnostic
LLMs, SLMs, your own
Architecture

A single container. Inside your network.

Everything ShareContext needs to plan, reason, govern, and remember runs inside one container deployed to your cloud, your VPC, or your Kubernetes cluster. It talks out to the warehouses, databases, and LLM gateways you already use. Nothing talks back.

ShareContext container · 1×
Engine · Frontend · Knowledge Graph · MCP Gateway
CNLA Engine
Plan · execute · assess · summarise
Skills + Methodology
Declarative YAML behaviours, governance-approved
Enterprise KG
Apache Jena Fuseki + OpenSearch index
Decision & Evidence Store
Postgres — bundled or your own
MCP Gateway
JSON-RPC tool plane, plug your own MCPs in
Web UI
React SPA · served by nginx in-container
Reads from · authorised connectors
BigQuery
Snowflake
Postgres
Redshift
!
Zero data egress. All warehouse traffic stays inside your network. LLM calls route through your gateway — Bedrock, Vertex, Azure OpenAI, or a private model — under your contract.
Engine

Skills, governed planning, co-authored metrics.

Frontier LLMs decide what to do. The execution plane decides whether — and how. Capabilities are declarative, auditable, and improve every time the system runs.

— 01

AI-drafted, governed planning

The planner is physically incapable of referencing entities outside the allowed-schema whitelist. Every generated query is statically validated before it touches data. Plans are inspectable; users approve at the right gates.

  • Plans approved before execution — at the user, team, or policy level
  • Recovery + retry built in: when stuck, the agent asks rather than assumes
  • Every step replayable from the evidence ledger
— 02

Skills as institutional IP

Each skill is a deterministic, governance-approved agent behaviour written in declarative YAML — vetted SQL patterns, segmentation logic, methodology recipes. Skills are versioned, diff-able, and promoted via a Draft → Approved workflow.

  • Same input → same output, every time
  • No engine release to ship a new skill
  • Skills compose — chained by agents to answer multi-step questions
— 03

Co-authored metrics

Metrics are not someone's BI dashboard. They're proposals from AI and analysts, vetted by governance, and committed to a shared definition. Metric Studio captures formula, owner, refresh cadence, and the trend points behind each value.

  • AI drafts a metric from a question — humans approve
  • One definition. One value. One source of truth across the org
  • Edit the formula → the engine re-runs and re-records the lineage
— 04

Context scanners

Point a scanner at a warehouse and it auto-extracts the schema, samples shapes, infers entity types, and proposes a starting Knowledge Graph. Reviewers approve nodes/edges in minutes — the graph compounds with every analysis after that.

  • Schema, joins, and cardinalities ingested automatically
  • Human-in-the-loop confirmation before anything goes live
  • The graph gets richer with every run — methodology stays in the system
Time to value

From kickoff to first answer in minutes.

No multi-month integration. The container is the integration. Run it, point it at a warehouse, and watch the engine produce value the same day.

Minute 0
Container deployed
One docker run (or a single Helm chart) on your cloud or on-prem cluster. The container brings its own Postgres, Fuseki, OpenSearch, and nginx — or wires up to your existing ones via env vars.
Minute 5
Identity + first warehouse connected
Plug in your IdP (OIDC) and a service account for the warehouse you want to start with. Read-only is enough.
Minute 10
Context scanner discovers your data
The scanner walks the warehouse, proposes a starting schema map, and seeds the Knowledge Graph. A reviewer approves the entities that matter.
Hour 1
First strategic query answered
Ask anything about your business, data, strategy, or process. The engine plans, fetches, validates, and commits an Executive Synthesis with the load-bearing figures up front.
Day 1
First metric, first decision, first goal
Co-authored metrics live in Metric Studio. The first decision is captured with its rationale. Goals committed with success criteria show live progress.
Week 1
Methodology starts compounding
Skills get drafted from successful runs. The graph thickens. Reasoning gets faster, cheaper, and more accurate every cycle — the system gets smarter the longer it runs.
LLM Routing

The system learns which model to use when.

No vendor lock-in. ShareContext is model-agnostic — Gemini, Claude, GPT, open-weights, on-prem SLMs. The engine auto-selects per step, learns from outcomes, and respects whatever policy you set on top.

Workflow step
Default selection
Plan + decomposition
Strong reasoner (Gemini 2.5 Pro / Claude Opus)
SQL generation
Fast generalist (Gemini 2.5 Flash)
Validation / dry-run
Local SLM or static rules
Disambiguation Q&A
Fast generalist
Executive synthesis
Strong reasoner

Policy-aware. Cost-aware.

Pin a model per step, set per-user or per-org budgets, route by data sensitivity. Sensitive workloads can stay on a private endpoint while everything else goes to the cheapest competent model. The engine learns the trade-off automatically from observed accuracy + cost.

Per-user API keys are supported for cost attribution. Or use a single corporate gateway — the call surface is identical.

Modularity

Every capability is an MCP.

ShareContext is built on the Model Context Protocol — every tool, connector, and skill is a first-class MCP. Add your own. Scale them independently. Govern them at the policy layer.

— Plug-in

Drop in your own MCPs

Internal APIs, custom data sources, line-of-business tools, third-party services — anything that speaks MCP is a first-class peer to the built-in capabilities. No engine release required.

— Scale

Replicate independently

Heavy connectors run as separate processes/containers. The engine talks JSON-RPC over the MCP gateway, so any single MCP can be sharded, replicated, or moved to a private region without touching the rest of the stack.

Top questions

What CIOs and investors ask first.

Q01 Where does the data live, and what crosses the perimeter?

Nothing crosses the perimeter that you didn't approve. The container runs inside your VPC. Warehouse traffic stays internal. LLM traffic goes through your gateway (Bedrock, Vertex, Azure OpenAI, or a self-hosted model) under your contract and audit log.

The Decision Ledger, Knowledge Graph, methodology, and skills are your data — stored in Postgres + Fuseki inside the container. Yours to keep, to export, to migrate.

Q02 How is this different from a copilot or a chat-over-data tool?

Copilots draft text and reach for tools. ShareContext is a control plane. It plans against an allowed-schema whitelist, executes through deterministic skills, gates at user/policy/governance checkpoints, records evidence for every step, and learns from outcomes.

Copilots help one person. ShareContext makes the organization's reasoning repeatable and replayable.

Q03 Will it scale?

The container starts at 4 vCPU / 16 GB and scales horizontally. Heavy MCPs run as separate processes — replicated, sharded, or moved to dedicated nodes without redeploying the engine. The Knowledge Graph and Decision Ledger sit on the same Postgres/Fuseki you'd scale for any read-heavy workload.

For multi-region or multi-tenant deployments, the same container is the building block — wire them together with your service mesh.

Q04 Vendor lock-in? What about lifting and shifting out?

The data is yours. Skills are declarative YAML. Methodology lives in your KG. The Decision Ledger is portable Postgres. The Knowledge Graph is standard RDF/SPARQL via Apache Jena. None of it is a proprietary blob.

Model-side: the engine speaks every major frontier model and is model-agnostic by design. Migrate between Gemini, Claude, GPT, or a private model without touching skills.

Q05 Do we need to migrate data into ShareContext?

No. ShareContext reads your warehouses where they are. The context scanner indexes metadata into the KG (table names, columns, joins, cardinalities) — not the rows themselves. Row-level data is fetched on demand, with an audit trail, and stays in your warehouse.

Q06 How does the engine pick which LLM to use?

Each workflow step (Plan, SQL gen, Validate, Synthesize, Disambig) has a default model class — strong-reasoner vs. fast-generalist vs. local SLM. The engine learns from accuracy + cost telemetry which model is winning for which step, and shifts traffic accordingly.

Operators can override at any granularity: per step, per skill, per user, per data sensitivity tag. Policy beats learning.

Q07 What does deployment actually look like on day one?

One container image. Pull it into ECR/GCR/ACR. Run it on Cloud Run, Fargate, GKE, EKS, OpenShift, or a single VM — anywhere a container runs.

Two env vars are mandatory: your IdP issuer and a warehouse connection string. Everything else is optional. Postgres is bundled; swap it for your own with a single env var.

Q08 How is governance enforced?

Three layers. Schema: the planner cannot reference entities outside the allowed-schema whitelist. Skills: only Approved skills run in production; Draft skills are sandboxed. Policy: pause-and-ask gates fire at user-defined sensitivity boundaries.

Every step, prompt, query, and intermediate result is logged. Compliance teams replay any decision from the ledger.

Q09 What gives this a moat?

The Knowledge Graph + Methodology library compound per-customer. The system learns the customer's data, their conventions, their decision patterns — and that learning becomes their asset, not ours. Frontier models change weekly; the methodology that knows your business doesn't.

Models commoditise. Context doesn't.

Q10 What about audit, SOC2, and inherited compliance?

Because the container runs in your VPC, it inherits the controls of your cloud — network policy, KMS, IAM, audit log, encryption-at-rest. SOC2 / ISO27001 evidence for the deployment is whatever your platform team already produces. ShareContext adds an in-app audit ledger (every prompt, every step, every approval), exportable to your SIEM.

One container away from a system of context.

Talk to us about your deployment. Design partner cohort open now.