Capabilities
Connect Enterprise Knowledge. Trace the Evidence. Govern the Intelligence.
kiLM combines governed model adaptation, live enterprise federation, knowledge and trace graphs, multimodal retrieval, compliance controls and approval-gated agents inside a dedicated single-tenant deployment.
Build & Adapt Domain AI
Adapt supported open-weight models to your domain and govern how they're evaluated, promoted and served — entirely in your boundary.
Fine-Tuning
kiLM turns your governed corpus into training-ready cohorts and fine-tunes open-weight models with LoRA / QLoRA (built on Unsloth and TRL's supervised trainers) — entirely on your own hardware, inside your boundary.
- LoRA / QLoRA supervised fine-tuning on your own governed data.
- Runs on your hardware, GPU-gated and air-gap friendly — nothing is sent out.
- Adapters are versioned, signed and eligible for evaluation and controlled promotion. Automatic evaluation depends on its gate, dataset and operating configuration, and real training requires the correctly built and validated GPU worker.
- Built entirely on permissive open source (Unsloth, TRL).
- Adapt supported open-weight models to your approved internal terminology, APIs and coding conventions, so assistants reason about your own programs.
Reinforcement Learning
kiLM closes the feedback loop. The ratings and preferences your users generate in governed chat become preference datasets, and kiLM aligns models to your team's judgement — in the same in-boundary pipeline.
- Preference optimisation (DPO / ORPO / KTO) to align models to your team's judgement.
- Learns from real human feedback captured in governed chat.
- Preference-tuned adapters are versioned and selectable per task.
- Same in-boundary, governed pipeline as fine-tuning — no data egress.
Domain-Specific SLMs
kiLM distills large teacher models into small, domain-specialised language models trained on your own governed corpus, so a compact model that runs cheaply still captures your domain. Distillation is in Preview.
- Knowledge distillation from a governed teacher into a compact, task-fit student model.
- Trained only on your own corpus — no data leaves your boundary.
- Runs on modest CPU / single-GPU hardware for low-cost, high-throughput inference.
- Promoted only after a comparative eval beats the baseline; versioned and selectable per task.
Task-Aware Model Selection & Routing
kiLM recommends and selects from models that are actually installed, licensed and compatible with the task and available hardware. Administrators can preview, approve or override the task policy within the vetted shortlist. Production-scale concurrent serving of many bases and adapters remains on the roadmap.
- Selects a best-fit model per request by task, hardware and cost, from installed and licensed models.
- Hardware-aware — respects the RAM and GPU your install actually has.
- Admin preview shows the routing decision before any policy goes live.
- Per-task overrides, including your own fine-tuned adapters.
Connect & Structure Enterprise Data
Reach the systems you already run and give their data a governed, typed shape — without rip-and-replace ingestion.
See it in a use case
MCP — AI Context Gateway
kiLM speaks the Model Context Protocol both ways. As a gateway it exposes your governed knowledge — documents, metrics, the graph and related entities — to MCP-aware clients; as a client it pulls grounded, permissioned context from external MCP sources.
- Connect Claude Desktop, Cursor or Cline to governed kiLM context.
- Per-user tokens, admin-managed, honouring your access policy.
- Two-way MCP — serve your context and pull from external MCP sources.
- Answers stay grounded and cited; no data is duplicated out of kiLM.
Federated Queries
Ask questions across the enterprise systems you already run — without copying data into kiLM — and get answers at query time. Optional governed caching and typed materialisation can be enabled for repeated workloads. Live SQL, REST and MCP sources require runtime network access.
- Query authorised operational systems in place — no extract, no duplication.
- RDBMS and data warehouses: native connection or the generic JDBC / ODBC path.
- Per-column sensitivity classification travels with federated results.
- Same role and access policy as native retrieval — one governance model.
Ontology & Schema Governance
kiLM reads your unstructured corpus and proposes a typed ontology and schema — entities, fields and relationships — through a visual editor and a review-and-approve proposal lifecycle, then projects approved models into a typed store you can query and report on.
- Proposes entities, fields and relationships from your raw documents.
- A visual ontology editor with a review-and-approve proposal lifecycle.
- Approved schemas project into a typed store you can query and report on.
- Schema drift and migration proposals surface as your corpus grows.
Reason, Trace & Verify
Turn connected data into evidence-grounded answers, traceable reasoning and verifiable compliance.
See it in a use case
Chat and Image / 3D Shape Search
Diagrams, scans, screenshots and photos are first-class content in kiLM. A vision encoder embeds images so you can search them by what they actually show, and 3D geometric similarity finds CAD parts by shape.
- Search images by content — diagrams, scans, screenshots, photos.
- Multimodal retrieval pairs the right figure with the text that explains it.
- Visual results honour the same sensitivity and access rules as text.
- Runs in-boundary on your own hardware — no third-party vision API.
- 3D shape search — find CAD data by geometry, surfacing parts and assemblies with a similar shape (not just by name or metadata).
Compliance Verification
kiLM treats compliance as something you can verify, not just assert. Once obligations, policies and evidence mappings are configured, kiLM continuously evaluates governed evidence against the standards and internal SOPs you're held to, surfaces findings and tracks remediation.
- Verify design and manufacturing processes against the standards and regulations you're held to (ISO, IEC and sector-specific), with cited evidence — once obligations are configured.
- Validate work against your company-specific process compliance — encode your internal SOPs, standards and gates, and check work products against them.
- Policy-as-code (OPA) enforces access and data-handling rules at runtime.
- Subject-erasure, audit trail and sensitivity enforcement built in.
Evidence-Grounded Answers & Assurance
kiLM retrieves permissioned evidence, returns inspectable citations, applies configured quality and contradiction checks, and can abstain when evidence is insufficient.
- Answers are grounded in retrieved, permissioned content and returned with citations.
- Configured quality and contradiction checks can score answers against their sources.
- Low-evidence questions get an honest "not enough information" — not a guess.
- Sensitivity and access rules apply before anything is cited back.
Build Governed Agents
Compose agents from governed templates and let them act across systems — proposed, policy-checked, approved and audited.
Visual Agent Studio & Governed Actions
Compose agents visually from governed templates with typed inputs and structured outputs, then run them under tool allowlists, budgets and an evaluation-and-approval lifecycle. Write-back actions to connected systems are approval-gated with auditable execution traces — Preview: recorded execution today; live adapters not yet shipped.
- Visual Agent Studio with a governed template library.
- Typed inputs and structured outputs.
- Tool allowlists and budget limits per agent.
- Evaluation and approval lifecycle before promotion.
- Approval-gated write-back actions — Preview: recorded execution today; live adapters not yet shipped.
- Auditable execution traces for every run.
Time-Series / Operational Data
Streaming operational signals — SCADA logs, sensor measurements and regular-cadence industrial telemetry — are handled by POLYCRACY together with its partner companies, outside the kiLM platform.
- Delivered by POLYCRACY and its partner companies — outside the kiLM platform.
- SCADA logs, sensor measurements and regular-cadence industrial telemetry.
- Machine- and deep-learning algorithm development, real-time performance analytics and functional digital twins on distributed systems.
Design the Capability You Need
Tell us the systems, models and governed outcome you need. We'll identify the required modules, hardware, connectors and live-validation plan.