Agent prototypes

Four agents GTDataWorks is building as working demonstrations and internal tools. Each one maps to a service in the catalog, and each ships with the same receipts a client build would.

The four prototypes

These are prototypes and specifications, not shipped products. The intake agent exists today as the brief builder on the contact page: a static form with no backend and no model call. The other three are specified below to the level a build starts from. Nothing on this page calls a live model.

PrototypeWhat it doesStatusEstimate

Customer Intake Agent

Collects a project brief from a visitor and turns it into a leadLive as the brief builder on the Contact page2 weeks to full CRM-connected versionDetails

Data Expert Showcase Agent

Answers questions over a document corpus and cites the passages it usedPrototype spec with scripted walkthrough3–4 weeksDetails

Lead Qualification Voice/SMS Agent

Calls or texts new leads, asks the qualifying questions, books demosPrototype spec4 weeksDetails

Internal Knowledge Base Agent

Answers questions about internal documents and code, within the asker's accessPrototype spec2–3 weeksDetails

Status is literal. Live means the static brief builder on the contact page, which calls no model. Prototype spec means the design on this page, with an estimate in working weeks for a scoped build; a scoping call turns an estimate into a quote.

Customer Intake Agent

Qualifies a visitor, gathers the project brief and hands a complete lead to the studio, with a receipt of the conversation attached.

Flow: from visit to lead

The CRM write, the booking, the notification and the receipt are the two-week build. The live brief builder stops after the conversation: it assembles the brief for you to send, with no backend and no model call.

Purpose

Automate the first contact without hiding anything from the person on the other end. The agent asks what you are trying to build, collects budget band, timeline and contact details, answers the usual pre-sales questions from a fixed set of studio facts, and books a follow-up call. Today the form of record is the brief builder on the contact page: a static, no-backend version of the same flow that composes the brief for you to send.

Core features

  • Conversational form-filling: name, company, project type, budget band, timeline, contact details
  • Pre-chat FAQ answered only from the published catalog: prices, timelines, what ships
  • Lead scoring against the catalog: service fit, budget band, urgency
  • Lead created in HubSpot or Salesforce through the REST API, with the brief attached
  • Receipt per session: what was asked, what was answered, what was written to the CRM, and when

Integrations

HubSpot or Salesforce REST API to create the lead. n8n or Zapier to send the confirmation email. Google Calendar API to book the follow-up call. Hosted on Modal or Google Cloud Run, with the model behind the OpenAI or Anthropic API.

Data Expert Showcase Agent

A public assistant that answers questions over a document corpus and shows which passages the answer came from.

Sequence: one question, end to end

The model only sees the retrieved passages, so every sentence in the answer can be traced to a document id on the receipt.

Scripted walkthrough

Scripted walkthrough, not a live model

Fixed text that steps through one query the way the showcase agent would handle it: question, retrieval, the passages it found, the answer with its sources, and the receipt. Nothing is generated when you click. The three documents are a fictional sample corpus.

Step 1 of 5
  1. Visitor
    Summarize Q3 sales data.
  2. Retrieval

    Embedded the question and searched the sample index: 3 documents, 41 chunks. 3 passages scored above the 0.70 cut-off.

    index sample-corpus, top-k 5, cut-off 0.70
  3. Top documents
    1. [1] q3-sales-summary.md, section 2, score 0.84Q3 revenue closed at $1.24M, 8% above Q2.
    2. [2] regional-breakdown.csv, rows 12–40, score 0.79Midwest 61% of new orders, West 22%, Northeast 17%.
    3. [3] board-notes-october.md, section 4, score 0.71September includes a one-off wholesale order of $90k.
  4. Agent

    Q3 revenue was $1.24M, up 8% on Q2 [1]. Most of the gain came from the Midwest, which took 61% of new orders [2]. September includes a one-off wholesale order of $90k; without it, Q3 is level with Q2 [3].

    Revenue by quarter, $M. Sample data.

    Sources
    1. q3-sales-summary.md, section 2 (0.84)
    2. regional-breakdown.csv, rows 12–40 (0.79)
    3. board-notes-october.md, section 4 (0.71)
  5. Receipt
    Receipt written: the question, the 3 document ids with scores, the model and prompt version, a hash of the response, and the timestamp. The reply above carries the receipt id, so the answer can be re-checked later against the same passages.

Purpose

Show retrieval-augmented generation working the way a client build would. The question goes to an index, the top passages come back with scores, the model answers from those passages and nothing else, and the reply lists its sources. A visitor can inspect the whole chain instead of taking the answer on trust. The public corpus is a small fixture of sample documents, not client data.

Core features

  • Question answering over a document corpus with embeddings and top-k retrieval
  • A citation on every claim: document, section and retrieval score
  • Summaries of long documents on request
  • A basic chart from retrieved tables on request
  • Receipt per query: question, document ids and scores, model and prompt version, response hash

Integrations

Embedded on the site as a React widget. Session logs go to a Postgres table for the metrics in the sidebar. Optionally the GitHub API, so the same agent can answer questions about a public repository as a technical demo.

Lead Qualification Voice/SMS Agent

Calls or texts a new inbound lead, delivers a short approved pitch, asks the qualifying questions and books a demo when the answers fit.

Flow: from new lead to booked demo

Every outbound call or text passes a policy check first, and every contact ends in a receipt whether or not the lead qualifies.

Purpose

For B2B clients with more inbound leads than hours to call them. The agent reaches each new lead within minutes over Twilio, by voice with SMS as the fallback, follows a script you approve, records the answers, and either books a demo or hands the lead back with notes. Consent and opt-out are handled at the start of every contact and written to the receipt. This is the hardest of the four to build: voice adds latency, interruptions and transcription quality to the test plan, which is why the estimate is four weeks.

Core features

  • Outbound calls through Twilio Voice, with speech recognition and synthesis on the call; SMS conversation when the call is not answered
  • Pitch and qualifying questions approved by you and versioned, so a change in the script shows up in the receipts
  • Opt-out honoured on the first ask and recorded
  • Demo booked through the Google Calendar API, invite sent to the lead
  • CRM updated with the outcome, the answers and the receipt id

Integrations

HubSpot or Salesforce REST API for the lead record and the write-back. Twilio Voice and Messaging. Google Calendar API for the booking, with a Zoom link as an option. Orchestrated in Node.js with a Python backend; each outbound action is gated by a policy check and logged as a JSONL receipt, the Terminus Gate Protocol pattern.

Internal Knowledge Base Agent

Answers questions about internal documents and code, scoped to what the person asking is allowed to read.

Flow: a question, checked and answered

Ingestion runs separately: a scheduled script embeds documents, wiki pages and repositories into the index and labels each chunk with its source and access group.

Purpose

Built for the studio's own use first: find the LangChain example from a past project, summarize the SOP on agent security, list which repositories use a given library. It is not client-facing, but it is the same pattern as a client's internal knowledge agent, and it is where the access-control and logging work gets proven before it ships to anyone else.

Core features

  • Retrieval over documents, wiki pages and repositories, with per-source access labels
  • Access check on every question: read-only, scoped by OAuth to the sources the asker may see
  • Answers cite file and section, and say “not in the sources” rather than guess
  • Exposed as an MCP server, so Claude, ChatGPT or any MCP client can use it as a tool
  • Receipt per question: who asked, which sources were searched, which were returned

Integrations

Slack or a command-line client for questions. GitHub for repositories. A scheduled ingestion script that records what it indexed and when. Served as an MCP server on Modal or Google Cloud Run.

Build one for your team

Pick the prototype closest to what you need and name the systems it has to touch. The estimates above are working weeks for a scoped build; a scoping call turns one into a fixed quote.