Case study
How Progranza builds and runs its systems with Claude
By Shyam Sembagoundar, Founder and Principal Architect · Published November 14, 2026
Progranza is a founder-led Salesforce consultancy that uses Claude in three ways. We use Claude Code to build and administer Salesforce orgs, for ourselves and for clients. We run the Claude API and Agent SDK inside our own marketing and outreach system. And we run Claude models on Amazon Bedrock inside a clinical documentation platform. This page shows what each looks like in practice, what is live today, and what is not.
At a glance
| System | How Claude is used | Status today |
|---|---|---|
| Growth Engine (marketing and outreach) | Claude API and Agent SDK, 13 specialist agents, built with parallel Claude Code sessions | In use internally for our own brands |
| Our internal Salesforce org | Claude Code with the Salesforce CLI to build and update objects, records and outreach journeys | In use internally |
| A U.S. pharmacy client's Salesforce and Marketing Cloud | Claude Code for development, setup, integration design and feasibility analysis | Client delivery. Claude is the builder, not part of their product. |
| Nurvivo clinical documentation | Claude Sonnet and Haiku on Amazon Bedrock in 15 backend functions | Pre-production |
How we work with Claude
- Claude Code is our build environment. Our Salesforce projects are Salesforce DX repositories with Salesforce agent skills installed, plus the Salesforce CLI and mcdev for Marketing Cloud, so Claude retrieves and deploys through the same tools our team uses.
- Every project keeps a CLAUDE.md and a memory index, so a new session starts with the connection details, conventions and past decisions already in context.
- Larger builds run as parallel sessions in isolated git worktrees, one workstream per session.
- Hooks make Claude state what a file affects before it can edit it, and we review what ships.
Example 1: Growth Engine, a multi-agent marketing and outreach system
We run three brands and wanted audits, strategy, content and outreach to be systematic instead of ad hoc. We built Growth Engine, a web app where agents find what is missing for a brand, turn gaps into simple tasks for a person, and draft content that a human approves before anything is published.
What we built
- 13 specialist agents, including a chief, content strategist, competitor researcher, site SEO and GEO auditors, a producer and a QA gate. Each is a markdown file that names its exact Claude model, its allowed tools, a per-invocation budget cap and a structured output schema.
- One prompt, two runtimes. The same file is loaded by Claude Code during development and by the production worker through the Claude Agent SDK.
- An independent QA gate. A fresh-context agent ties every claim in a draft to an approved fact, scores voice and flags compliance issues, then returns a structured pass or fail before a person sees the draft.
- One metered door for model calls. Web code can only call Claude through a single wrapper that resolves the workspace's key and writes a usage ledger row for every call, with tokens and cost.
- Versioned writing rules. Trials write two versions side by side, a per-brand experiment flag controls who gets the new version, and a rollback action restores the previous one.
- Outreach starts only from targets a person has approved. Approved targets then move through people search, enrichment and research.
- The code uses three Claude tiers, Opus 5.5, Sonnet 5.5 and Haiku 4.5, with each agent definition naming its own model.
How it was built
Five workstreams ran as parallel Claude Code sessions, each in its own worktree and branch: sending and email events, contacts and consent, campaigns and the send engine, the Salesforce connector, and journeys. Later outreach phases, LinkedIn and research and autopilot, followed the same pattern.
Status: in use internally for our own brands, with leads syncing to our internal Salesforce org. It is not offered to customers today.
Example 2: our internal Salesforce org
Our growth operation runs in our own Salesforce org. Claude Code builds and updates it with the Salesforce CLI: custom objects, outreach journeys, and a Growth app that tracks the proposals we submit on Upwork, with status and Connects spent. When a proposal goes out, Claude checks the live listing first, then updates the record.
Status: in use internally.
Example 3: Salesforce and Marketing Cloud for a U.S. pharmacy client
On client delivery, we use Claude Code to build the Salesforce and Marketing Cloud applications and custom integrations, working inside our normal Salesforce workflow. Three moments show how we use it.
- Marketing Cloud connected once. We connected the client's Marketing Cloud through mcdev, so any new session opened in that project can retrieve and deploy without re-authenticating, and the setup history sits in project memory.
- A tested result beat model knowledge. For the integration with the client's pharmacy order platform, Claude proposed receiving inbound webhooks through Experience Cloud. Our team had already tried that and Salesforce rejected it, so the design shifted toward middleware that relays into Salesforce through a Connected App with an integration user.
- Verify before promising. When the client asked whether an assigned sales rep could be pulled from the platform, Claude searched the platform's API schema first. It found no such field, noticed that the existing order feed already carries a sales person per order, and recommended two checks before we said yes.
Status: client delivery. Claude is the builder here. It is not part of the client's product.
Example 4: Nurvivo, clinical documentation on Amazon Bedrock
Nurvivo's nurse practitioner platform calls Claude through Amazon Bedrock. Claude Sonnet 4.6 is the default chat model and Claude Haiku 4.5 the default fast model, both configurable by admins. Fifteen backend functions use it, covering transcription processing, clinical note formatting, structured clinical data extraction, medication extraction and drug-interaction checks, ICD-10 and CPT code suggestions, treatment-plan extraction, pre-visit summaries, patient letters and post-op instructions.
- Admin tools to draft, preview and publish prompt configurations and to update guardrail settings, so prompt changes are controlled and reviewable.
- Per-call token and cost logging by function and model.
- AI drafts, a clinician signs. Clinical notes go through a sign-off step before they are final.
- Built under HIPAA and PHIPA requirements, with a design rule that PHI stays inside the system boundary.
Status: pre-production.
What we do not claim
Claude is not part of any customer-facing product we sell today. We use it to build and run our own systems and to deliver client Salesforce work. Where a system is internal or pre-production, this page says so. We checked every statement here against our code and project records on November 14, 2026.
What we learned
- Verify against the source before promising a client anything. Reading the API schema took minutes and changed the answer.
- When a tested result contradicts the model, the tested result wins.
- Route every model call through one metered wrapper from the start. Cost and usage then stay visible instead of being reconstructed later.
- Use an independent reviewer with fresh context rather than asking the drafting agent to check itself.
- Keep prompts as files in the repository, so the version Claude Code develops against is the version production runs.
Where this goes next
We are formalizing a Claude implementation practice for customers: Claude with Salesforce integrations, agentic automation, and Claude Code adoption for engineering teams. The examples above are the working methods we bring to it. If you are weighing something similar, a 15-minute call with the architect is the quickest way to scope it.