AI integration for SATX companies that operate under real regulation.
Veteran-owned, SDVOSB-certified, San Antonio-based. We deploy Claude, GPT-4, and open-weight LLMs against insurance, healthcare, defense, and professional services workflows — with the compliance posture those industries actually require.
Most AI pilots in San Antonio die between the demo and production.
The pattern repeats across SATX. A carrier in the USAA orbit runs a slick GPT demo on synthetic claims data, then discovers their actual claim files are PDFs with handwriting, and Guidewire doesn't expose the fields the prompt assumes. A clinic in the Medical Center wants chart summarization but no one signed a BAA. A JBSA-adjacent contractor wants RAG over CUI but the vendor only deploys to public cloud. A 40-attorney firm pays for Copilot licenses nobody opens. The technology works. The integration with regulated data, real systems of record, and actual user workflows is where it falls apart.
- ▸ Demos run on clean data; production data is messy PDFs, handwritten notes, and 20-year-old mainframe extracts
- ▸ Compliance scoping happens after the model is picked, forcing expensive rebuilds when BAAs or FedRAMP authorization don't exist
- ▸ Vendors disappear after the pilot, leaving brittle prompts and zero documentation for your IT team
- ▸ Offshore engineers cannot get on-site with underwriters, clinicians, or cleared facility staff — and that is where requirements actually live
We deploy AI against your real data, not a sandbox demo.
- STEP-01
Workflow audit, not vision deck
We sit with your underwriters, coders, paralegals, or claims adjusters for two to five days. We map the exact decisions, the systems of record (Guidewire, Epic, NetDocuments, QuickBooks), and where an LLM actually saves minutes per task.
- STEP-02
Data boundary and compliance scoping
Before any model touches data we lock down the boundary: HIPAA BAAs, PII classification, FedRAMP-aligned hosting if needed, on-prem or Azure Government for cleared work. We pick Claude, GPT-4 class, or open-weight Llama based on what the data allows.
- STEP-03
RAG pipeline over your corpus
We build retrieval over your policies, claim files, EHR notes, contracts, or SOPs using pgvector or Azure AI Search. Chunking and metadata are tuned per document type. Citations are mandatory in every response so reviewers can verify.
- STEP-04
Human-in-the-loop rollout
First release goes to a single team with shadow mode: the model drafts, humans approve. We instrument acceptance rate, edit distance, and time saved. We tune prompts and retrieval weekly until the numbers hold, then expand.
- STEP-05
Handoff with runbooks
You get the source code, the eval harness, the prompt library, and a runbook your IT team can operate. We stay on retainer for model upgrades and drift, but you are not locked into us. No black boxes.
# RAG pattern we deploy for HIPAA-bounded clinical Q&A
# Runs in Azure (BAA-covered) or on-prem with Llama 3 + pgvector
from anthropic import AnthropicBedrock
from app.retrieval import search_corpus
from app.audit import log_phi_access
client = AnthropicBedrock(aws_region="us-gov-west-1")
SYSTEM = """You answer clinical workflow questions for credentialed staff.
Rules:
- Cite every factual claim with [doc_id:section].
- If retrieved context is insufficient, say so. Do not infer PHI.
- Never output a patient identifier that was not in the question.
"""
def answer(user_id: str, question: str, patient_mrn: str | None):
chunks = search_corpus(
query=question,
filters={"acl": user_id, "mrn": patient_mrn},
k=8,
)
log_phi_access(user_id, patient_mrn, [c.doc_id for c in chunks])
context = "\n\n".join(f"[{c.doc_id}:{c.section}]\n{c.text}" for c in chunks)
resp = client.messages.create(
model="anthropic.claude-3-5-sonnet-20241022-v2:0",
max_tokens=1024,
system=SYSTEM,
messages=[{"role": "user", "content": f"{context}\n\nQ: {question}"}],
)
return resp.content[0].textReference RAG skeleton for a regulated deployment: ACL-aware retrieval, mandatory citations, PHI access logging, and a model hosted under a BAA.
Field FAQ.
→ We're an insurance carrier in the USAA orbit. Can you actually integrate an LLM with Guidewire or our claims platform?
Yes. Most of the carrier work we do touches Guidewire ClaimCenter, Duck Creek, or a homegrown mainframe-backed claims system. We integrate through documented APIs where they exist and through message queues or batch extracts where they don't. Typical first wins are first-notice-of-loss summarization, coverage determination drafting, and subrogation letter generation. We keep the model out of the system of record — it drafts, the adjuster commits. That avoids a class of audit problems.
→ How do you handle HIPAA when deploying LLMs for South Texas Medical Center clients?
We deploy under signed BAAs with the model provider — Anthropic via AWS Bedrock, Azure OpenAI, or self-hosted Llama on infrastructure you own. PHI never transits a non-covered service. We log every retrieval that touches a chart, enforce ACLs at the vector store level so a clinician only retrieves on patients they're assigned, and run de-identification on training and eval data. The compliance posture is reviewed with your privacy officer before any pilot goes live.
→ We're a JBSA contractor working with CUI and some classified-adjacent data. Can you support that?
This is one of the reasons our SDVOSB status matters. We work in IL4 and IL5 environments via Azure Government and AWS GovCloud, and we have engineers who hold or can obtain clearances. For air-gapped work we deploy open-weight models — Llama 3.1 70B or Mixtral — on customer-owned hardware so no inference traffic leaves the enclave. We will not pretend FedRAMP Moderate is good enough for IL5 data. Scope drives the architecture.
→ Why hire a San Antonio firm instead of a cheaper offshore AI shop?
Two reasons that matter in practice. First, regulated AI work fails when the engineers can't sit with the underwriter, nurse, or paralegal whose job is changing — and they need to be in the room, not on a 10pm Zoom from twelve time zones away. Second, federal and defense-adjacent work has citizenship and location requirements offshore vendors cannot meet. For SATX clients, we drive to your office. That is not a small thing.
→ What does a typical AI integration engagement cost and how long does it take?
A scoped pilot — one workflow, one team, real data, measurable outcome — usually runs eight to twelve weeks and lands in the $80K to $180K range depending on data complexity and compliance scope. Production rollout across additional teams is typically a follow-on engagement or a monthly retainer. We do not sell six-figure discovery phases. If we cannot describe the deliverable in one paragraph, we will not take your money.
→ We're a mid-size San Antonio law or accounting firm. Is AI integration worth it at our scale?
Often yes, but not always for the reasons vendors pitch. The wins for professional services firms are usually document review acceleration, first-draft generation for routine filings or memos, and intake triage. We've seen 30 to 60 percent time reduction on specific document tasks in pilots. We won't sell you a chatbot. If your bottleneck is partner review capacity, an LLM that drafts faster is leverage. If your bottleneck is rainmaking, it isn't.
→ Will you train a custom model on our data, or just use ChatGPT?
Almost always the right answer is retrieval-augmented generation against a frontier model, not fine-tuning. Fine-tuning is expensive, brittle to model upgrades, and rarely beats good retrieval plus careful prompting. We fine-tune when there is a specific output format the base model gets wrong consistently, or when latency or cost requires a smaller self-hosted model. We will tell you which case you're in before you spend money on the wrong one.
→ What happens after the pilot? Are we locked into your platform?
No. You own the code, the prompts, the eval harness, and the infrastructure-as-code. Everything runs in your cloud account or on your hardware. We use standard tools — LangChain or direct SDK calls, pgvector or Azure AI Search, GitHub Actions for CI. If you want to take it in-house after handoff, your team can. Most clients keep us on a light retainer for model upgrades and drift monitoring, but that's a choice, not a contract trap.
→ How do you measure whether the AI integration is actually working?
Before we ship anything we agree on three numbers: a quality metric (acceptance rate, edit distance from final, or task-specific accuracy on a held-out eval set), a time metric (minutes saved per task, measured against a pre-pilot baseline), and a guardrail metric (hallucination rate, PHI leakage incidents, or policy violations). We instrument all three from day one. If the numbers don't move, we kill the feature. We've killed features.
Continue recon.
AI integration services
How we scope, build, and ship LLM workflows against regulated enterprise data.
REL-02Selected case studies
RAG and workflow automation projects we've shipped for regulated industries.
REL-03Pilot packages
Fixed-scope AI pilots: eight to twelve weeks, one workflow, measurable outcome.
REL-04Talk to an engineer
Skip the SDR funnel. First call is with someone who will write the code.
Ready to put a real LLM against your real data? Let's meet in San Antonio.
Talk to a VooStack operator. We respond within one business day.