AI & Machine Learning Company in San Antonio

We quote in ranges tied to scope. A focused feasibility and baseline study, meaning data profiling, leakage checks, a benchmark against your current process and a written go or no-go recommendation, typically runs USD 25,000 to 55,000 over four to eight weeks and regularly saves clients far more than it costs by stopping a project that was never going to work. A single production model with a full pipeline, meaning ingestion, feature engineering, training, evaluation, serving, monitoring and documentation, typically runs USD 90,000 to 250,000 depending on how messy the source data is and how many systems it has to touch. A platform engagement covering a feature store, several models, a retraining pipeline, an experiment-tracking setup and governance artifacts runs USD 250,000 to 700,000 and up across phases. Computer vision on a plant floor sits at the higher end because data collection, lighting rigs and edge deployment are real engineering rather than a notebook. San Antonio rates sit well below San Francisco and New York at equivalent seniority, and below Austin as well. GPU and cloud spend is billed at cost and estimated up front so it is not a surprise line item.

Both took effect September 1 2025 and both change architecture. SB 1188 imposes duties on covered entities handling electronic health records. It requires that EHRs of Texas residents be physically maintained in the United States, with the deadline for that localization requirement set at January 1 2026, and it requires access be limited to individuals who need the record to perform treatment, payment or health care operations duties, with appropriate administrative, physical and technical safeguards. It also permits licensed practitioners to use AI in diagnosis and treatment planning provided they act within their scope of licensure, review the AI-generated records according to Texas Medical Board standards, and disclose the diagnostic use of AI to the patient. SB 815 works from the payer side: a utilization review agent may not use an automated decision system to make, wholly or partly, an adverse determination. Automated systems are still allowed for administrative support and fraud detection, and the commissioner may audit and inspect a utilization review agent's use of an automated decision system at any time. The Act took effect September 1 2025 and its requirements reach health benefit plans delivered, issued or renewed on or after January 1 2026. So we build these models as decision support with human sign-off, keep training and inference in United States infrastructure, and log the practitioner AI-use disclosure SB 1188 calls for as part of the pipeline.

Explainability here is not a research preference, it is a paperwork requirement. Under ECOA and Regulation B a declined applicant is entitled to specific principal reasons for the adverse action, and under FCRA the same applies where a consumer report contributed. Texas Department of Insurance Bulletin B-0003-26, issued June 12 2026, adds the expectation that regulated entities maintain robust governance, risk-management controls and internal audit functions for AI systems, encourages verification and testing methods that identify errors and bias, and applies to any third party working with the regulated entity. Practically, we design for reason codes from the start rather than bolting them on. That means preferring gradient-boosted trees or generalized additive models with monotonic constraints on variables where direction of effect should be obvious, computing SHAP values at inference and storing them next to the decision, mapping SHAP contributions to a fixed adverse-action reason taxonomy your compliance team approves, and running disparate-impact testing across protected classes on every retrain with results written into the model card. When a deep model genuinely outperforms, we keep it behind a simpler challenger and document the tradeoff instead of hiding it. The output your examiner sees is a model card, a testing record, and a decision log.

Yes, and the hard parts are almost never the model. San Antonio industrial buyers, whether that is Toyota Motor Manufacturing Texas and its supplier park, an aerospace MRO operation at Port San Antonio, or an engine test cell at a research organization, run into the same four problems. Lighting varies across shifts and the model trained on morning images fails at night, so we specify fixed lighting and capture a full-cycle dataset before we train. Defect classes are wildly imbalanced because good parts vastly outnumber bad ones, so we lean on anomaly detection and synthetic augmentation rather than pretending we have a balanced classification problem. Labels come from operators who disagree with each other, so we measure inter-rater agreement first and fix the label guide before training. And inference has to happen locally because a network hiccup cannot stop a line, so we quantize and export to ONNX Runtime or TensorRT and run on an industrial PC or Jetson-class device with a local buffer. We instrument confidence thresholds so uncertain parts route to a human rather than being guessed, and we track false-negative rate as the metric that matters, because a missed defect costs far more than a false alarm.

ERCOT is its own world and modeling for it does not transfer cleanly from other markets. Load and price forecasting works on fifteen-minute settlement intervals, so the target granularity is fixed for you. Weather is the dominant driver and Texas weather has fat tails, which is why models trained only on typical years fell apart during Winter Storm Uri in February 2021 and why we insist on holdout periods that include extreme events even when they hurt the headline metric. Distributed solar and battery adoption keeps shifting the net-load shape, so features have to include behind-the-meter estimates and the model needs a shorter retraining cadence than most forecasting problems. For a municipally owned utility like CPS Energy the operational uses are load forecasting, outage prediction from asset age and weather, vegetation-management prioritization, and non-technical loss detection on AMI data. For refining and midstream operators such as Valero and NuStar the work looks different: yield optimization, turnaround planning, corrosion and equipment-health modeling from historian data, and demand forecasting on product movements. In both cases the model has to be paired with a written operating procedure describing what a human does when it is wrong, because it will be wrong on exactly the days that matter most.

The Texas Data Privacy and Security Act took effect July 1 2024 and it reaches further than most state privacy laws because it does not use revenue or record-count thresholds the way California's does; it applies to entities conducting business in Texas and processing personal data, with a genuine small-business carve-out. Three provisions shape ML pipelines. Sensitive data, which includes precise geolocation, biometric identifiers, health data and children's data, requires opt-in consent before processing, so those columns get a consent gate before they reach a training set. Universal opt-out signals such as Global Privacy Control must be honored, which means the opt-out has to propagate into feature stores and downstream training corpora, not just the marketing database. And deletion, correction and portability rights follow the data, so we build deletion propagation across the warehouse, the feature store, cached embeddings and any fine-tuning corpus, then test it in the QA suite rather than assuming it works. The Attorney General enforces exclusively with a cure period that does not sunset, and Texas has shown it will pursue large privacy matters, including a 1.4 billion dollar biometric settlement with Meta and a 1.375 billion dollar settlement with Google. Data residency defaults to Azure South Central US in San Antonio unless a specific requirement moves it.