AI Development in the US

About AI Development

Most AI projects fail in the same place: a demo works on ten hand-picked examples, then falls apart on real user input, real data volume, and a real invoice at the end of the month. We build the unglamorous parts that decide whether an AI feature survives contact with production: retrieval that returns the right context, evaluation you can run before every deploy, guardrails on what the model is allowed to do, and cost tracking per request so the bill is a number you chose rather than one you discover.

What the engagement includes

• Retrieval-augmented generation over your own documents and databases
• AI agents with tool access, scoped so they can only do what you allow
• Model Context Protocol (MCP) servers that connect assistants to your systems
• LLM integration into existing products: OpenAI, Anthropic, and open models
• Document extraction and classification for invoices, contracts and forms
• Evaluation suites so a prompt change cannot quietly break output quality
• Cost and token controls with per-request tracking and hard budget limits
• Chatbots and support assistants grounded in your own content, not the open web

What FTC model deletion, California training-data law and ADMT rules ask of a US AI build

The remedy that sets an American AI build apart is that the Federal Trade Commission can order the model itself destroyed. In Everalbum, settled on 11 January 2021, a photo app that had used customers' pictures to build face recognition without the consent it claimed had to delete any facial recognition models or algorithms developed with those photos and videos, and to get express consent before using biometric information to create face embeddings. In Rite Aid, announced on 19 December 2023, the order banned the retailer from using facial recognition for five years and required it to delete, and to direct third parties to delete, the images collected and any algorithms or other products developed from them. The failures the FTC listed there read as an engineering checklist: no testing or documentation of accuracy before deployment, low-quality images let into the system, no ongoing monitoring of the false positive rate, and staff never trained that the tool could be wrong. The commercial lesson is that a model is only as durable as the least defensible dataset that went into it, and an order of this kind follows the data into every product trained on it, including products a vendor built. So on a US engagement we keep a lineage record that ties each model version and each fine-tuned adapter to the datasets, consents and licences behind it, which means a disputed source can be pulled and only the affected versions retrained, rather than the whole system written off. Evaluation runs before release on data that looks like production, including the blurred, partial and badly formatted inputs real users send, and error rates are tracked by segment after launch, with a named person who reads them rather than a dashboard nobody opens. Whether a given dataset was collected with the consent your intended use needs is for your own counsel.

California now makes training data a published document. Civil Code section 3111, enacted as AB 2013, requires the developer of a generative AI system or service made available to Californians to post documentation about the data used to train it, on or before 1 January 2026 for systems released since 1 January 2022, and before each new or substantially modified release after that. Developer includes anyone who substantially modifies a system for use by the public, and section 3110 says a substantial modification includes the results of retraining or fine-tuning. So a company that fine-tunes an open-weight or hosted model and puts the result in front of the public is plausibly the developer of a new release, with its own page to publish. The required list is specific: the sources or owners of the datasets, the number of data points, which may be given in ranges, whether they contain copyrighted, trademarked or patented material, whether they were purchased or licensed, whether they include personal information or aggregate consumer information, any cleaning or processing, the collection period, and whether synthetic data was used. Systems whose sole purpose is security and integrity, aircraft operation, or national security work available only to federal entities are excluded. The California AI Transparency Act sits beside it for generators of image, video or audio with over 1,000,000 monthly visitors or users. From 2 August 2026 such a provider must embed a latent disclosure carrying its name, the system's name and version, the time of creation and a unique identifier, offer users a visible disclosure, and provide a free detection tool, with penalties of $5,000 per violation and each day a separate violation. Large online platforms take on duties from 1 January 2027. We keep a dataset manifest from the first experiment, because this document cannot be reconstructed from memory at launch, and treat provenance marking as a stage of the generation pipeline. Whether you are a developer or a covered provider is for your own counsel.

If the model decides something about a person, two US regimes decide how it ships. California's privacy regulations, approved on 22 September 2025, govern automated decisionmaking technology used for a significant decision, meaning one that provides or denies financial or lending services, housing, education enrollment or opportunities, employment or contracting opportunities or compensation, or healthcare services. A business already using such a system has until 1 January 2027, and anything deployed after that date has to comply from its first use. That means a pre-use notice explaining the specific purpose in plain language at or before the point the data is collected, a right to opt out, and a right of access that covers the logic of the system, the output for that person and how it was used in the decision. The rules treat a tool as replacing human decisionmaking unless a reviewer knows how to interpret its output, actually analyzes it alongside other relevant information and has authority to change the result, so a recruiter clicking approve on a ranked list does not take a tool out of scope. A human appeal with those same qualities is one way around the opt-out. New York City has enforced Local Law 144 since 5 July 2023: an employer using an automated employment decision tool to screen candidates or employees in the city needs an independent bias audit from the previous year, calculating selection or scoring rates and impact ratios across sex, race and ethnicity and their intersections, a published summary, and notice to candidates ten business days ahead. The city's guidance says the audit uses the tool's historical data, test data only where that is insufficient, and that a vendor may commission an audit while the employer stays responsible. So the system logs each input, output and reviewer action per decision, can produce a per-person explanation, and exports outcomes by category in the form an auditor works from. Which of these reaches your product is for your own counsel.

How we work with US clients

A US engagement gets agreed twice. Once with the person who wants the thing built, and then again with legal, security, procurement and finance, who arrive separately and rarely in the same week. That second round is where a supplier outside the country loses the calendar: a master services agreement to redline, a security questionnaire, a certificate of insurance, and a vendor portal that assumes everyone it onboards can sign a W-9. Send all of it while the scope is still being written rather than in the week you wanted to start, and where the honest answer to a row is that we do not hold something, we write that in rather than leaving it blank.

You are also buying from a company you cannot drop in on, so the evidence has to do the work an office would otherwise do. What we send is live production URLs rather than a deck, a written scope with one fixed number against it, and the names of the people who will do the work rather than a bench that changes after signature. What we will also tell you is the shape of what you are hiring: a small team in Noida that has been building since 2021, not a firm that can put four more people on your project next Monday. If that is what your timeline needs, we say so before the contract rather than after it.

Indicative pricing in USD

Converted from our published Indian rates and rounded. Exchange rates move, so the figure on your quote is calculated on the day it is issued.

Questions US clients ask

AI Development: common questions

Other markets we work in