OKLAHOMA CITY · AI INTEGRATION

AI integration for Tinker, the Aeronautical Center, and OKC's energy and insurance core

Veteran-owned, SDVOSB-certified. We deploy Claude, GPT, and self-hosted models inside regulated boundaries — CUI, PHI, ITAR — with the same-day on-site support OKC actually expects.

Veteran-Owned SDVOSB
[001 / 005] Field Conditions

Most AI pilots in OKC die between the demo and the SCIF

// SITUATION

We see the same pattern across Tinker sustainment shops, FAA software vendors at the Aeronautical Center, energy operators along Robinson, and the insurance carriers downtown. A vendor runs a slick demo on synthetic data. Then someone asks where the data actually lives, whether the model trains on prompts, how it handles CUI or PHI, and what happens when GPT-4o gets deprecated mid-contract. The demo team has no answers. The pilot stalls. Twelve months later the workflow is still being done by hand in Excel, Maximo, Guidewire, or WellView — and the budget is gone.

  • ▸ Vendor demos on public APIs that cannot legally touch CUI, ITAR tech data, or PHI claims notes
  • ▸ RAG systems built without citations, so depot engineers and adjusters cannot trust or verify outputs
  • ▸ No eval harness, so a model upgrade silently breaks production extraction the next quarter
  • ▸ Offshore delivery teams that cannot badge into Tinker, Will Rogers, or a HIPAA-controlled facility
SDVOSB
Certified veteran-owned, federally eligible
Same-day
On-site response across the OKC metro
6–10 wks
Typical pilot to measured production value
[002 / 005] Operational Approach

Ship AI that survives audits, depot reality, and an actual production load

  1. STEP-01

    Workflow audit before models

    We sit with depot planners, claims adjusters, or land techs for a week. We map where humans retype data between SAP, Maximo, Guidewire, or WellView. AI goes where the rekeying happens, not where the demo looks good.

  2. STEP-02

    Data boundary and classification

    Before any prompt leaves a VPC we classify what touches CUI, PII, PHI, or proprietary log curves. We pick model hosting accordingly: Azure GovCloud, Bedrock, on-prem Llama, or vendor APIs with signed BAAs and zero-retention terms.

  3. STEP-03

    RAG with citations or it ships nothing

    Every answer points back to the source paragraph in the tech order, policy doc, well file, or claim note. Adjusters and depot engineers will not trust an LLM that cannot show its work, and we will not deploy one that does not.

  4. STEP-04

    Evals tied to real tickets

    We pull 200–500 historical cases — depot work orders, FAA software defect reports, denied claims — and run them as a regression suite on every model swap. No vibes-based prompt engineering, no silent drift across model versions.

  5. STEP-05

    On-site rollout and handoff

    Same-day drives to Tinker, the Aeronautical Center, or downtown energy towers. We pair with your engineers in their SCIF or office, document runbooks in your repo, and leave the system operable without us.

// PYTHON PATTERN 
# RAG pipeline used on a depot tech-order assistant at an OKC sustainment shop.
# Constraint: tech orders are CUI//SP-PROP. Nothing leaves the gov tenant.

from voostack.rag import Retriever, Answerer, EvalHarness
from voostack.guards import PIIScrub, CitationRequired, DomainAllowlist

retriever = Retriever(
    index="tinker-techorders-v4",        # chunked TOs + IPB + AFTO 781 history
    embed_model="bge-large-en",          # local, no egress
    top_k=8,
    rerank="bge-reranker-v2",
)

answerer = Answerer(
    model="azure-gov:gpt-4o",            # IL5-eligible tenant
    system="You are a depot maintenance assistant. Cite TO paragraph for every claim.",
    guards=[
        PIIScrub(),                      # strips tail numbers, SSNs, names
        CitationRequired(min_citations=1),
        DomainAllowlist(["techorders", "ipb", "afto781"]),
    ],
    temperature=0.1,
)

# Regression suite: 312 historical work orders with known correct TO references.
harness = EvalHarness("workorders_2019_2023.jsonl")
report = harness.run(retriever, answerer)

assert report.citation_accuracy > 0.95
assert report.hallucination_rate < 0.02
report.publish("s3://voostack-evals/tinker-to-assistant/")

Trimmed-down RAG harness from a Tinker-area sustainment engagement. Citations and evals are non-optional.

[003 / 005] Common Questions

Field FAQ.

→ Can you handle CUI or ITAR data on AI projects tied to Tinker AFB work?

Yes. We deploy LLM workloads inside Azure Government, AWS GovCloud, or on customer-controlled GPUs depending on the data classification. For CUI//SP-PROP and ITAR-marked tech data we keep embeddings, prompts, and completions inside the accredited boundary and use models with signed zero-retention terms or self-hosted Llama, Mistral, or Qwen weights. We are SDVOSB and our staff hold active clearances where the contract requires it.

→ What does AI integration actually look like for an FAA software vendor at the Mike Monroney Aeronautical Center?

Usually it is internal: pulling answers out of 8000-page system requirements documents, summarizing defect reports across Jira and ServiceNow, or generating first-draft test procedures from requirements. We do not touch operational ATC paths. We treat FAA AVS and AJM documentation as the source of truth, build retrieval over it, and require citations so engineers can verify against the controlling document before anything ships.

→ We are an OKC oil and gas operator. What AI work actually pays back here?

Three things consistently return value: extracting structured data from scanned well files and historical logs, surfacing lease and division-order anomalies before they become title problems, and natural-language access to production and SCADA history so engineers stop opening five tools to answer one question. We integrate with WellView, Quorum, Enverus, and OSDU rather than building parallel data lakes.

→ Insurance carriers and TPAs in OKC are nervous about LLMs touching PHI and claims data. How do you handle that?

We sign BAAs, deploy in HIPAA-eligible environments, and route PHI only to models with no-training and zero-retention contractual terms — or to self-hosted models on your infrastructure. We log every prompt and completion for audit. For claims summarization and adjuster copilots we build human-in-the-loop interfaces so a licensed adjuster signs off before any decision affects a policyholder.

→ How fast can you actually be on-site in Oklahoma City?

Same day for anything inside the metro and Tinker, Will Rogers, or the Aeronautical Center. Our engineers are based in the U.S. and we keep staff who can drive or fly into OKC on short notice. For ongoing engagements we co-locate weekly. We do not do offshore handoffs and we do not pretend remote-only works for accredited facilities where badges and SCIFs are involved.

→ What does a typical AI integration engagement cost and how long does it take?

A scoped pilot — one workflow, real users, measured against a baseline — typically runs 6 to 10 weeks and lands in the mid five figures to low six figures depending on data integration scope. Production rollouts with eval harnesses, monitoring, and handoff documentation usually run 3 to 6 months. We price fixed-fee for pilots and time-and-materials or labor-category for longer federal work.

→ Do you build with Claude, GPT, or open models? How do you decide?

All three, decided by data classification and task. GPT-4o and Claude Sonnet are our defaults for commercial work with strong reasoning needs. For CUI or ITAR we use GovCloud-hosted GPT or self-hosted Llama 3.1 70B and Qwen 2.5. For high-volume extraction we often run smaller fine-tuned models. We do not marry a vendor — the eval harness tells us which model wins on your data.

→ Can you augment our existing engineering team instead of running the whole project?

Yes. Staff augmentation is a large part of what we do. We place senior engineers — typically 12+ years of experience, U.S.-based, clearable — embedded inside your team using your tooling and process. They commit to your repo, attend your standups, and are accountable to your tech lead. This is common for OKC clients who have the domain expertise but need AI and modernization muscle for a defined window.

→ Why does SDVOSB status matter for an AI project?

For commercial buyers it does not, beyond signaling that we work to federal-grade documentation and security standards. For federal buyers it matters a lot: SDVOSB sole-source authority goes up to $7.5M for non-manufacturing under the VA and similar thresholds elsewhere, and SDVOSB set-asides at Tinker and other DoD activities let contracting officers move faster than open competition. We hold the certification and we are comfortable with the paperwork.

[004 / 005] Adjacent Files

Continue recon.

REL-01

AI integration services

How we scope, build, and hand off RAG and LLM workflows.

 REL-02

Engagement case studies

Sustainment, energy, and regulated-data AI rollouts we have shipped.

 REL-03

Pilot packages

Fixed-fee 6–10 week AI pilots scoped to one real workflow.

 REL-04

Talk to an engineer

Skip the discovery call. Bring the actual problem.
[ NEXT ACTION ]

Have a real AI problem in OKC? Let's talk to your engineers, not your marketing team.

Talk to a VooStack operator. We respond within one business day.

Open a Channel Full Services