INDEA
Sovereign AI — engineered to stay on your terms.
INDEA is the indigenous AI platform for on-prem inference, retrieval-augmented reasoning, and mission-tuned fine-tuning over classified data. Model weights, training data, and operational telemetry never leave customer infrastructure.
Inference
On-Prem · Air-Gapped
Models
Customer-Owned
Data
Never Leaves Site
Stack
GPU · CPU · Edge
Inference, reasoning, retrieval — on your terms.
INDEA is built around three load-bearing capabilities. Each is deployable on its own; together they form a sovereign reasoning surface for the operational data you cannot send elsewhere.
Sovereign Inference
Open-weight foundation models or your own fine-tunes — running on your hardware.
Run open-weight foundation models or customer-tuned variants entirely on-prem. No call-home, no remote telemetry, no hosted dependencies. Inference happens on hardware you control, under audit you own.
Throughput, latency, and quantisation profiles are tuned per deployment — we don't ship a one-size-fits-all binary.
- 01GPU, CPU, and edge inference targets
- 02Quantisation, batching, and KV-cache tuning per deployment
- 03Deterministic outputs for reproducible evaluations
- 04Tamper-evident inference logs
Mission-Tuned Reasoning
Fine-tune on your own data — without sending a byte to anyone else.
Fine-tune foundation models on classified data without third-party access. Training, evaluation, and weight management stay on customer infrastructure. Resulting weights are bonded to your hardware.
Evaluation harnesses run on your own benchmark sets, so model quality is measured against your operational tasks — not someone else's.
- 01On-prem fine-tuning and LoRA workflows
- 02Customer-owned evaluation harness
- 03Weight provenance and lineage tracking
- 04Reversible roll-back across model versions
Retrieval Over Classified Corpora
Retrieval that respects your classification handling.
Build retrieval indexes over classified documents and operational corpora. Per-document and per-section classification flows through retrieval, into prompts, and into the final response.
Need-to-know enforcement is applied at retrieval time, not as a post-hoc filter. The model never sees content the user isn't authorised to see.
- 01Per-document and per-section classification labels
- 02Need-to-know enforcement at retrieval time
- 03Citation-grounded responses with source linkage
- 04Hybrid dense + sparse retrieval over operational data
The model goes to the data — never the other way around.
AI shouldn't change the threat model of the data it reasons over. INDEA's deployment posture is engineered around that constraint.
Air-Gapped
Runs without internet.
Models, embeddings, tooling, and updates are delivered offline via signed bundles. INDEA never assumes a network it doesn't own.
Customer-Owned Weights
Fine-tuned weights stay in your hardware.
Trained models are bonded to customer hardware and key material. No third-party model registry, no shared hosting tier.
Classified-Aware
Classification flows end-to-end.
Per-document and per-section classification labels travel through retrieval into prompts and final responses. The model never sees what the user can't.
Hardware-Agnostic
NVIDIA, AMD, CPU — whatever fits the mission.
Inference and fine-tuning across NVIDIA, AMD, and CPU-only targets. Edge profiles for forward-deployed nodes with limited compute.
How operators are using INDEA.
Anonymised examples drawn from active engagements. Formation and unit identifiers are illustrative.
01
Intelligence Cell
Retrieval-augmented reasoning over classified intel corpora.
Analysts query operational intel with citation-grounded responses and per-document classification handling on every result.
02
Operations HQ
Mission-tuned summarisation of multi-source reports.
Fine-tuned models summarise multi-source operational reports against the unit's own reporting conventions, not a generic style.
03
Technical R&D
Internal knowledge retrieval over decades of design docs.
Engineering teams query historical design documents, decisions, and trial data — without exposing intellectual property to a third party.
Other platforms in the operational spine.
OneCommand
(MapPlan)
Terrain, signals, logistics, and operations on one geospatial surface — GIS, HF/VHF/UHF propagation, fleet tracking, war planning, and surveillance.
Learn MoreATMA
End-to-end training lifecycle management — curricula, assessments, readiness, and skills analytics.
Learn MoreHEART
Enterprise administration spine — HRMS, ERP, analytics, CRM and helpdesk, and learning management — on one platform.
Learn MoreSovereign AI evaluated on your data, on your hardware.
We run proof-of-value evaluations on customer premises against customer benchmarks. No exfiltration, no shared corpora, no telemetry — only the model and your data, in one place.