Skip to main content
AI SoftwareCorporation
AI SoftwareCorporation

We build, modernize and support the software growing organizations run on.

+1 302-464-5943Eastern Time (ET)

Start a project

Generative AI & LLMsGenerative AI & LLM Solutions

Generative AI that cites its sources and asks before it acts

We build LLM-⁠powered assistants, agents and features into your products and tools, scored on real cases from your business before release.

Start a projectSee the approach
Hands hold two printed pages of diagrams up in front of an open laptop on a desk.

Outcomes

Answers with sources, actions with sign-⁠off

  • Answers your staff can check

    Assistants answer from passages retrieved from your approved content and link each claim to its source. When the sources hold no answer, they say so instead of guessing.

  • Agents that ask before they act

    Agents work through a short list of approved tools with scoped permissions. Refunds, record changes, customer messages and other consequential steps wait for a person to approve them.

  • Sensitive data kept in bounds

    Data use is agreed in writing before any model sees it. Where possible, models run inside your cloud account, and personal data a task does not need is masked first.

  • Quality and cost you can see

    Every change to a prompt, model or retrieval step is scored on your own test cases before release, and quality, latency and cost per request are tracked in production.

Capabilities

Grounded search, agents and guardrails

Most builds draw on several of these at once. An assistant needs retrieval, evaluation and guardrails, and an agent adds approved tools and sign-⁠off steps on top.

Grounded assistants and copilots

Assistants that answer from your policies, manuals, tickets and records through retrieval-⁠augmented generation (RAG), cite the passages they used and respect each user’s existing access rights.

  • Hybrid keyword and vector search with re-⁠ranking
  • Chunking and metadata tuned to your documents
  • Document permissions enforced at query time
  • Delivered in Microsoft Teams, Slack, intranets or apps

Agentic workflows with approvals

Agents that plan multi-⁠step tasks and call your systems through approved tools, from gathering case details to preparing a refund, while people approve the steps that matter.

  • Tools exposed as typed APIs or MCP servers
  • Scoped, short-⁠lived credentials for each tool
  • Sign-⁠off for refunds, edits and customer messages
  • Step limits, timeouts and a full action log

LLM features inside your apps

Summarize, draft, classify and extract within the products and internal tools people already use, with structured output your code checks before acting on it.

  • Summaries of cases, calls and long documents
  • Drafts that people edit and approve
  • Classification and tagging of free text
  • JSON output validated against a schema

Model selection, cost and latency

We compare hosted and open-⁠weight models on your own test cases and choose on quality, cost, latency and data residency rather than leaderboard scores.

  • Fine-⁠tuning only where prompts and retrieval fall short
  • A gateway layer that keeps models swappable
  • Simple requests routed to smaller, faster models
  • Prompt caching, streaming and token budgets

Evaluation and observability

Test suites built with your subject experts score every change to prompts, models or retrieval before release, and tracing shows what the system does in production.

  • Golden datasets drawn from real cases
  • LLM graders calibrated against human reviewers
  • Regression gates in CI for every change
  • Traces of prompts, retrievals and tool calls

Guardrails, privacy and security

Controls that keep sensitive data inside agreed boundaries and outputs within policy, each one attacked with adversarial inputs before release.

  • Defenses against direct and indirect prompt injection
  • Personal data masked before model calls
  • Model access through your cloud account and region
  • Training and retention opt-⁠outs applied where offered

Approach

Evaluation first, prompts second

See our delivery process

  1. Step 1: Choose a use case worth building

    The starting point is a task, not a model: who does it today, how you would judge a good answer, what a wrong one costs and whether a language model is the right tool at all.

    Activities

    • Sit with the people who do the task today
    • Review the content, systems and access rules involved
    • Compare an LLM with search, rules or a simpler fix

    You receive

    • A scoped use case with risks and review points
    • A written recommendation, including where not to use AI

  2. Step 2: Build the evaluation set first

    Before prompts are written, we collect real questions and cases with your subject experts and agree on what a correct, safe and useful answer looks like.

    Activities

    • Gather typical, awkward and adversarial examples
    • Write grading rubrics with your subject experts
    • Set acceptance thresholds and name unacceptable failures

    You receive

    • A versioned evaluation set and scoring scripts you own

  3. Step 3: Prototype and compare

    Competing models, prompts and retrieval designs run on your content and are scored against the evaluation set, with cost and latency measured alongside quality.

    Activities

    • Compare hosted and open-⁠weight models on the same cases
    • Tune chunking, retrieval and prompts against the scores
    • Estimate cost per request and response times at expected volume

    You receive

    • A scored comparison and a recommended design
    • Running-⁠cost and latency estimates

  4. Step 4: Harden for production

    We add guardrails, permissions, approval steps and tracing, integrate with your identity provider and systems of record, and review data flows with your security team before a pilot group gets access.

    Activities

    • Add injection defenses, data masking and output checks
    • Integrate through your APIs, identity and logging
    • Red-⁠team the system and fix what breaks

    You receive

    • A production release for a pilot group
    • Threat model, data-⁠flow diagram and runbooks

  5. Step 5: Operate and improve

    In production we trace requests, sample and grade outputs, and turn user feedback and reviewer corrections into new test cases. The full suite runs again before any change ships.

    Activities

    • Review flagged answers and user feedback
    • Pin model versions and re-⁠test before any upgrade
    • Adjust routing and caching as usage grows

    You receive

    • Regular quality, cost and latency reports
    • A growing evaluation set and improvement backlog

Deliverables and fit

Evaluation sets, code and runbooks you keep

What you receive

8 deliverables
  • Use-⁠case brief with a written go/no-⁠go recommendation
  • Evaluation set, grading rubrics and scoring scripts in your repository
  • Model comparison covering quality, cost per request and response times
  • Production code for retrieval, prompts, tools and integrations
  • Guardrail, permission and data-⁠handling configuration, with a data-⁠flow diagram
  • Red-⁠team findings and the fixes they led to
  • Tracing and dashboards for quality, cost and latency
  • Runbooks, decision records and handover sessions with your team
Two people's hands over a printed bar chart, one pointing at a bar and the other holding a pen.
Evaluation scores, reviewed with your subject experts

A good fit if

  • A generative AI pilot impressed in a demo but never reached production
  • Staff lose time searching policies, manuals or past tickets for answers
  • You want an assistant or agent in your product that holds up with real customers
  • Security or legal won’t approve AI until they know where your data goes
  • You are choosing between hosted models and running open-⁠weight models yourself
  • Model costs or response times keep rising without a clear cause

Engagement models

A pilot, specialists or support in production

The models that usually suit this service. You can switch as the work changes.

Compare engagement models

  • Project delivery

    A scoped pilot or first release, judged against the evaluation set agreed with your experts.

  • Team extension

    LLM evaluation or retrieval specialists joining a team you already run.

  • Managed support

    Assistants and agents in production that need monitoring, re-testing and model upgrades.

Technology

Models, retrieval and evaluation tools

Listed so you can check the fit with your stack. None of them implies a partnership or certification.

Tools and platforms

  • Azure OpenAI
  • Amazon Bedrock
  • Google Vertex AI
  • Anthropic Claude
  • Llama
  • Hugging Face
  • vLLM
  • Model Context Protocol
  • LangGraph
  • pgvector
  • OpenSearch
  • Langfuse
  • Promptfoo
  • Microsoft Presidio

Illustrative scenario

Illustrative scenarioGathering incident evidence so on-⁠call engineers start from factsRead the scenarioHide the scenario
Illustrative scenario

Gathering incident evidence so on-⁠call engineers start from facts

The platform team at a software company, whose on-⁠call engineers respond to alerts across many services, dashboards and runbooks.

Challenge
The first stretch of every incident goes on opening dashboards, searching logs, checking recent deployments and finding the right runbook. Much of the know-⁠how sits with a few senior engineers, and incident notes are pieced together from memory afterward.
Approach
  1. 1Turn past incidents with known causes into an evaluation set, graded by senior engineers
  2. 2Give an agent read-⁠only tools for metrics, logs, traces, deployment history and runbooks
  3. 3Post a timeline and likely causes to the incident channel, each linked to the query or log line behind it
  4. 4Let the agent propose runbook steps such as a rollback, which run only after the on-⁠call engineer approves them
  5. 5Mask customer data before model calls, cap steps and spend per incident, and trace every tool call
Outcome
On-⁠call engineers start from a timeline with the evidence attached instead of a blank screen. Every suggestion links to the data behind it, fixes stay in human hands, and each change to the agent is replayed against past incidents before release.

Services involved

Ask about a project like this

FAQ

Questions to settle before an LLM goes live

Ask a question

Will our data be used to train someone else’s model?

Our practice is to use enterprise terms and settings that exclude your data from model training, and to record each provider’s current terms with you in writing, because terms change. Where we can, we reach models through your own cloud account in a region you choose, and switch off optional data retention when the provider allows it. When data must not leave infrastructure you control, we run open-⁠weight models there instead.

How do you stop an assistant from making things up?

Answers are grounded in passages retrieved from your approved sources, each one shows its citations, and the assistant is instructed to say it does not know when nothing relevant is found. We measure faithfulness on your own test questions before release and keep sampling answers in production. No method removes errors entirely, so topics where a wrong answer is costly are routed to a person.

Can AI agents take actions in our systems safely?

Within limits you set, yes. Agents call a short list of tools with scoped, short-⁠lived credentials, every call is logged, and consequential actions, such as issuing a refund, editing a record or messaging a customer, need a person’s approval first. We start with read-⁠only or easily reversed actions and widen what an agent may do only when evaluation results and review history support it.

How do you choose a model and keep running costs under control?

We test candidate models on your own cases and choose the smallest one that clears the quality bar, since cost and response time tend to grow with model size. Cost per request and latency are estimated during the prototype, then tracked in production against budgets and alerts. Caching, streaming and routing simple requests to smaller models help keep both in range as usage grows, and a gateway layer keeps a later change of provider contained.

When is generative AI the wrong tool?

When the result must be exact and repeatable, such as a price, a tax calculation or an eligibility rule, conventional code is cheaper, faster and easier to audit. It is also a poor fit when there is no reliable content to ground answers in, or when nobody can check the output and a mistake would be costly. In those cases we suggest deterministic code, better search or a change to the workflow itself, and we can build that instead.

Related services

Elsewhere in the stack

Next step

Where would an assistant or agent help first?

Describe the system, process or decision in front of you. Expect questions back, a sensible first step and a fitting engagement model.

First step
A conversation about the problem, your systems and constraints
You leave with
A view on whether we can help, and a sensible first step
Before any work
A written proposal covering scope, team, approach and terms
Commitment
None until you approve that proposal
Start a projectSee how we work

Or call +1 302-464-5943 · Eastern Time (ET)