Skip to content
Enterprise AI

Generative AI that cites its sources.

We build retrieval-augmented knowledge engines, private LLM deployments and autonomous agents for organisations where a wrong answer is expensive — grounded in your data and governed by your policies.

Request an AI auditExplore AI agents
knowledge.querygrounded

Which vendor contracts cap our indemnity below $5M?

Three active agreements cap indemnity below $5M. The largest exposure is Northwind Logistics, capped at $2M under the 2024 MSA[1], which the risk register flags as high-priority for renewal[2]. Legal guidance recommends renegotiating caps at renewal[3].

  • [1]Master Services Agreement v4.pdf§ 12.3 · p. 18
  • [2]Vendor risk register.xlsxRow 214
  • [3]Legal wiki / IndemnitiesUpdated Sep 2026
recall@5
99.1%
p95
84ms
unsupported
0
[01]What we build

Beyond chatbots. Systems that do the work.

Four building blocks, usually combined: knowledge to ground answers, agents to act on them, private models to keep data in, and automation to run it all day.

[01]

Knowledge engines

Hybrid vector and graph retrieval over your documents, tickets and databases, with every answer traced to its source.

  • Multi-modal parsing for tables and diagrams
  • Permission-aware retrieval
Learn more
[02]

Autonomous agents

Agents and multi-agent systems that plan, call your tools and verify each other before anything reaches production.

  • LangGraph state machines with checkpoints
  • Human approval on high-stakes actions
Learn more
[03]

Private model enclaves

Open-weights models such as Llama, DeepSeek and Mistral served inside your VPC or datacenter, fine-tuned on your domain.

  • vLLM and TensorRT-LLM GPU serving
  • LoRA fine-tuning on internal language
Learn more
[04]

Cognitive automation

Replace brittle RPA scripts with workflows that understand contracts, invoices and email — and know when to ask.

  • Document-to-ERP extraction
  • Exception queues with full context
Learn more
[02]RAG pipeline

Seven stages between a question and a trustworthy answer.

Most RAG prototypes stop at embed-and-search. Production systems need every stage engineered, measured and observable — this is the pipeline we deploy.

RAG solutions
  1. [01]

    Ingest

    Unstructured · OCR

    PDFs, wikis, tickets, tables and scans parsed with layout and permissions preserved.

  2. [02]

    Chunk

    Semantic · hierarchical

    Split on meaning and document structure, not character counts, with parent links kept.

  3. [03]

    Embed

    Dense + sparse

    Each chunk gets dense vectors and BM25 terms, plus entities synced to a knowledge graph.

  4. [04]

    Retrieve

    Hybrid · GraphRAG

    Keyword precision fused with vector recall; multi-hop graph traversal for related entities.

  5. [05]

    Rerank

    Cross-encoder

    A cross-encoder rescores candidates so only the most relevant evidence reaches the model.

  6. [06]

    Generate

    Routed LLM

    The cheapest model that clears your eval bar answers strictly from the retrieved context.

  7. [07]

    Cite

    Groundedness check

    Every claim is linked to its source passage; unsupported sentences are removed or flagged.

Retrieval Accuracy
99.1%
Search Response Time
<85ms
Hallucination Rate
<0.1%
[03]From answers to actions

When knowledge needs to act, agents take over.

The same retrieval layer powers agent swarms that plan, execute and validate multi-step work. Pick a scenario and watch it run.

How our agents work
Objective

Provision a high-availability Kubernetes cluster, configure an Istio service mesh, and deploy 14 microservices with zero downtime.

ObjectivePLPlannerREResearcherEXExecutorVAValidatorOutput
  • PlannerTask decomposition & DAG
  • ResearcherCloud topology & quota analysis
  • ExecutorTerraform & Helm dispatcher
  • ValidatorCanary health & chaos probes
agent-session · microservices-deploy
0/7
  1. $ jp-swarm run --objective "Autonomous cloud swarm"
  2. ▍
[04]The shift

Scripted software versus reasoning systems.

Comparison of legacy scripted systems with AI reasoning systems
DecisionsLegacyRigid rules that break on unexpected inputAIReasoning loops that replan on novel edge cases
RetrievalLegacyExact keyword matches that miss meaningAIHybrid vector + graph search with verified citations
CoordinationLegacyManual handoffs between fragmented toolsAIAgents executing across systems with approval gates
ImprovementLegacyWeeks of recoding and regression testingAIEvaluation-driven updates and feedback loops
ScaleLegacyBound by headcount and working hoursAIElastic, around the clock, with full audit trails
[05]AI services

8 ways we put AI to work.

Each is a full engagement — architecture, build, evaluation and operations — and most clients combine several.

[06]Guardrails & compliance

Governed by design, not by policy memo.

Your risk team reviews the architecture, not a promise. Controls are enforced in code and evidenced in logs from the first prototype.

  • SOC 2 Type II
  • ISO 27001
  • HIPAA
  • GDPR
  • EU AI Act ready
  • OWASP LLM Top 10
01PII redaction
Sensitive fields masked before prompts leave your boundary.
02Injection shielding
Inputs and retrieved content screened for prompt-injection attempts.
03Zero retention
Provider agreements and private hosting so data is never trained on.
04Evaluation harness
Groundedness, accuracy and policy tests run on every release.
05Audit trail
Every prompt, source and tool call logged and attributable.
06Model routing
Workloads pinned to approved models and regions by policy.
AI feasibility audit

Find the AI wins hiding in your data.

Our AI leads review your data estate, security constraints and highest-cost workflows, then return a prioritised roadmap with expected ROI and a prototype plan.

Request an AI auditAI case studies