Case Study
Price-to-Win, Air-Gapped
Self-hosted estimation intelligence under NIST and CMMC; proposals from four weeks to five days.
Client
Mid-size defence contractor and manufacturer
Market
Defence & advanced manufacturing
Scale
Complex hardware and software programmes
Duration
12+ months
Team
9 engineers
Engagement model
Managed Delivery
Delivery hub
London & Belgrade
Stack
Self-hosted Llama, Crew.ai agentic framework, RAG, Python, React
Capability
AI Engineering
Challenges
Challenges
Budget overruns
Programmes consistently exceeded estimates by 20% to 40%.
Slow bidding
Manual proposal preparation took weeks, costing bids or winning them unprofitably.
Compliance constraint
Defence data required NIST CSF and CMMC alignment, with no external inference.
The solution
Air-gapped architecture
A fully self-hosted deployment on open-weight Llama models, with no external calls.
Agentic estimation
Specialised agents (hardware, software, risk) collaborating on each estimate.
Historical knowledge retrieval
RAG over past programmes and industry standards for grounded estimates.
Price-to-win support
Strategic recommendations against historical bid outcomes and market patterns.
The honest section
What we would tell you before starting
Self-hosting cost more to run than a hosted API and the base model lagged the frontier on the hardest tasks. We benchmarked against a hosted baseline during scoping so the trade-off was a decision, not a surprise. For this compliance regime, self-hosting was the only viable answer.