Generative AI Company in Houston

We give fixed-fee proposals, not open T-and-M estimates, and scope each engagement rather than publishing bands. A scoped Houston generative AI proof of concept is a seven to eleven week engagement covering a corpus audit, a baseline RAG pipeline on AWS Bedrock or Azure OpenAI, a basic eval harness, and a hosted demo. A custom production GenAI system for a Halliburton-class upstream subsurface RAG, an MD-Anderson-class clinical documentation pilot, or a Chevron-class HSE incident triage agent adds domain-specific embeddings, hallucination detection trained on the client's corpus, HIPAA BAA wiring for TMC work, and OSDU-aware chunkers for subsurface work across a ten-to-fifteen-week engagement. Full production GenAI systems with self-hosted Llama 3 or Mistral on dedicated GPU clusters (typical for ITAR-controlled NASA-adjacent work or strict subsurface IP isolation) sit at the top of the range. Houston sits roughly mid-market US, slightly above Dallas, below SF and NYC. We give fixed-fee proposals, not open T-and-M estimates.

Yes. Subsurface RAG is our most-requested Houston upstream engagement, and the operational reality is more nuanced than generic enterprise RAG. We ingest LAS well logs, DLIS log files, WITSML real-time drilling data, mud logs, seismic interpretation reports, geological surveys, completion reports, HSE incident reports, and decades of petroleum-engineering literature. Domain-specific chunking respects log section boundaries, formation tops, and drilling-phase transitions rather than naive paragraph splits. Embeddings combine domain-tuned sentence transformers (fine-tuned on petroleum vocabulary using SPE OnePetro and the public USGS corpus where licensing permits) with generic text embeddings in a hybrid retrieval setup. The OSDU data ecosystem (the upstream-industry-standard data platform run through The Open Group with ExxonMobil, Shell, Chevron, Halliburton, SLB, Baker Hughes, IBM, Microsoft, AWS as members) increasingly defines the data backbone — we build OSDU-compliant ingestion where the client has adopted it. Hallucination is the catastrophic failure mode for any operational use, so we ship hallucination-detection trained on petroleum corpora and a confidence-thresholded response policy.

HIPAA Privacy Rule, Security Rule, and Breach Notification Rule apply to every TMC clinical generative AI engagement, and the Texas Medical Records Privacy Act (Texas Health and Safety Code Chapter 181) layers stricter Texas-specific consent and breach-notification rules on top. We maintain HIPAA BAAs with AWS Bedrock, Azure OpenAI Service, Anthropic Enterprise, and OpenAI Enterprise plans, route all inference through covered endpoints, run PHI de-identification through AWS Comprehend Medical or Microsoft Healthcare NLP before any prompt construction, and log every inference to an audit trail with patient-record-correlated retention. Epic Hyperspace and Cerner-now-Oracle-Health integration uses SMART on FHIR with OAuth scopes minimized to the specific use case. We do not run TMC clinical PHI through hosted Claude / GPT-4o / Gemini endpoints without an active BAA and the necessary contractual residency commitments. For research-adjacent GenAI we route through the client's IRB and produce a Common Rule-aligned documentation package.

Yes, with caveats. NASA Johnson Space Center mission operations, SpaceX Houston engineering, Aerojet Rocketdyne, KBR human-spaceflight contracts, and Lockheed Martin Houston space-program work all touch ITAR-controlled USML Category XV (spacecraft, including ground support equipment and software) and EAR 600-series dual-use technology. We route all engineering through US-person staffing on US soil, deploy generative AI inside AWS GovCloud US-East (Ohio) or US-West (Oregon) with FedRAMP High inheritance, use self-hosted Llama 3 70B or Mistral Large on dedicated H100 GPU clusters with no egress to public LLM APIs, and document export-control compliance in writing before kickoff. CMMC Level 2 readiness is included where the contract requires it. We will not run ITAR-controlled technical data through hosted Claude / GPT-4o / Gemini endpoints regardless of latency or capability advantages. We coordinate with the contractor's empowered official and export-control office at every milestone.

The Texas Data Privacy and Security Act (TDPSA, effective July 1, 2024) gives Texas residents access, deletion, correction, portability, and opt-out rights and triggers Data Protection Assessment requirements for high-risk processing including profiling that produces legal or similarly significant effects. The Texas AG enforces TDPSA, and biometric data (faceprints, voiceprints, fingerprints) gets stricter treatment under the older Texas CUBI statute. For Houston generative AI engagements we run a TDPSA scoping review during discovery, produce a DPA for high-risk use cases, wire consumer opt-out plumbing into agent state management, and route TDPSA-sensitive workloads to US-region cloud (Azure South Central US is the closest hyperscaler, in San Antonio). For employment, lending, and housing decisions adjacent generative AI we add adverse-impact testing and a human-oversight loop. TMC clinical work falls under HIPAA preemption where the laws conflict, but TDPSA still applies to non-clinical patient-experience and patient-communication tooling.

When a project requires genuine research (clinical NLP novel architectures, computational biology, scientific ML for subsurface modeling, formal verification of clinical LLM outputs) we scope collaborations with Rice University Computer Science, U of Houston, UT Health Biomedical Informatics, or Baylor College of Medicine rather than overselling in-house capability. Our core team handles applied engineering, MLOps, and productionisation, which is where most Houston GenAI projects actually stall. For standard work (RAG, fine-tuning, prompt-injection defenses on known architectures) no academic partner is needed. We will tell you up front which bucket your problem fits into. We have introduced Houston clients to TMC3 innovation district programs, Rice's Ken Kennedy Institute, and U of Houston's Hewlett Packard Enterprise Data Science Institute when industry-affiliate programs fit their roadmap and budget.

Our default cloud landing zones for Houston are AWS Bedrock in us-east-2 (Ohio, ~35ms from Houston, full Bedrock catalog), Azure OpenAI in South Central US (San Antonio, ~50mi west of Houston metro at sub-15ms RTT — typically the fastest commercial LLM endpoint for Houston buyers), and GCP Vertex AI in us-south1 (Dallas, ~25ms RTT). For TMC HIPAA-covered workloads we use only BAA-covered endpoints (Bedrock, Azure OpenAI Service with HIPAA-eligible status, Anthropic Enterprise, OpenAI Enterprise). For NASA JSC, SpaceX Houston, KBR, Aerojet Rocketdyne, and Lockheed Martin Houston ITAR-controlled work we route through AWS GovCloud US-East or US-West with FedRAMP High inheritance and US-person staffing. For self-hosted Llama 3 / Mistral on dedicated GPU clusters (typical for subsurface IP isolation at ExxonMobil, Chevron, Halliburton, SLB, Baker Hughes scope) we use AWS p5 (H100) or p5e (H200) instances in us-east-2 or GovCloud as workload requires. TDPSA does not impose hard residency on its own but we document data flows for the Texas AG.

A typical Houston subsurface RAG production deployment (Halliburton-class internal copilot over LAS well logs and drilling reports, or ExxonMobil-Spring-HQ-class HSE incident triage) takes fourteen to twenty weeks from kickoff to production. Week 1 to 3 is corpus audit, domain-specific chunking strategy, and OSDU readiness check. Week 4 to 10 is embedding pipeline, retrieval optimization, baseline LLM selection (Claude 3.5 Sonnet on Bedrock vs GPT-4o on Azure South Central US vs self-hosted Llama 3 70B), and hallucination-detection training. Week 11 to 16 is observability, integration into the operational system (PI System adjacency, Halliburton DecisionSpace, OSDU-compliant ingestion), and human-in-the-loop checkpoints. Week 17 to 20 is staged rollout. TMC clinical deployments add three to six weeks for HIPAA BAA wiring, Epic / Cerner-Oracle-Health integration, and IRB approval where research-adjacent. ITAR-controlled NASA-adjacent work adds four to eight weeks for AWS GovCloud setup and export-control documentation.