What our tools are built with, and where they run.
Every tool we build runs inside your environment, on parts your IT team already knows. Here is each part, what it does in your tool, and, where one exists, the test run or live product behind it.
It runs where your data is allowed to live
Machines you already own
After a readiness check, we deploy on your servers or workstations. Nothing leaves your network, and we connect only when you let us in for support.
A cloud account in your name
If you'd rather not run hardware, we deploy into a GPU cloud account you own, India-hosted if your data must stay in India (E2E Networks and Yotta are two such providers). You hold the account and the bill; we never hold your data.
A machine you buy through a partner
If your machines aren't enough, you buy a suitable one through a hardware partner. You own it, the warranty sits with the manufacturer, and we set it up. We don't resell hardware.
India-hosted when it must be: AWS in Mumbai and Hyderabad, Azure in Pune, Chennai, Mumbai and Hyderabad, Google Cloud in Mumbai and Delhi, and GPU accounts with E2E Networks or Yotta.
Every tool can run in sealed mode, refusing any outbound connection except to the model you approved. We test it on every release.
Where your data livesEvery cloud and hosting service we deploy into
Default is what we reach for first. Also is what we use when the job calls for it. In your stack is what we work inside when your company already runs it.
Cloud and hosting
- VercelDefault
Hosting for both studio sites and most Next.js projects
- Cloudflare (Workers, R2, D1, CDN, DNS)Default
Edge APIs, micro-apps, files and caching
ig-comment-dm on GitHub - Amazon Web ServicesDefault
Client cloud accounts on AWS
- AWS LambdaDefault
Event-driven functions and APIs
- Amazon SESDefault
Transactional email, including both studio sites' own
Trust - Amazon EC2Also
Servers and GPU instances in the client's account
- Google Cloud (Cloud Run, Compute Engine, Cloud Storage)Default
Client cloud accounts on Google; containers that scale to zero
- Microsoft AzureAlso
Client cloud accounts on Azure, including Foundry
- India cloud regionsDefault
Data kept in India: AWS Mumbai and Hyderabad; Azure Pune, Chennai, Mumbai and Hyderabad; Google Cloud Mumbai and Delhi
- E2E Networks and YottaAlso
India-hosted GPU accounts in the client's name, for private models
- Your own machinesDefault
Office desktops, GPU workstations and plant servers we install onto
The smallest model that passes your tests
We try two or three models on 20 to 50 of your own examples and keep the smallest one that passes. Where a tool calculates, the numbers are worked out in code; the model only writes the words around them.
On your machines
Open-weight models on hardware you own, or buy through a partner.
- Platforms
- Ollama and LM Studio to try models; serving software, such as llama.cpp or vLLM, chosen per machine after the readiness check.
- When we choose it
- Data can't leave the network: charts, loan files, plant SOPs.
In your cloud account
The model service inside your own AWS, Google Cloud or Azure account, in the region you pick.
- Platforms
- Amazon Bedrock and Bedrock AgentCore; Gemini Enterprise Agent Platform; Microsoft Foundry; GPU accounts with E2E Networks or Yotta.
- When we choose it
- You want hosted models without the data leaving your account, or leaving India.
Direct, under contract
The model maker's own service, on terms that exclude training, with a DPA or BAA where needed.
- Platforms
- Claude API, OpenAI API, Gemini API.
- When we choose it
- Low volume, or a task that needs the strongest models, which can't run on your own hardware.
How we choose a model
- Start from 20 to 50 of your own examples, as the readiness check does.
- Try two or three candidates, local and hosted, and keep the smallest that passes.
- Check the licence. Nothing licensed for non-commercial use ships in client work.
- Wherever a tool calculates, the numbers come from code and the model writes only the words around them.
- Re-run your test set before any model update reaches your staff.
The AI parts inside a tool
- Reading documents into fields
- Reads a document into fields, leaves a field empty when the document is silent, and keeps the page and line.
- Text recognition first for scans, then a model held to a fixed set of fields; the Extractor engine.
- Search with citations
- Answers from your documents and quotes the version it used.
- Postgres with pgvector, plus keyword search.
- Routing with a person in the loop
- Drafts replies and decides which ones a person sees first.
- The Gatekeeper engine, with rules checked in code before any model call.
- Speech
- Transcribes calls and field conversations in Telugu, Hindi, English and other languages.
- Speech-to-text on the phone or an office machine, licence-checked.
- Portals with no API
- Fills and reads web portals at a pace that doesn't get the account locked.
- Playwright, paced and logged by the Reconciler.
- Agents and MCP
- Lets an AI tool act in your systems under your sign-in and your rules.
- MCP servers we build or configure; Claude Agent SDK, OpenAI Agents SDK or Microsoft Agent Framework, matching your platform.
- Testing
- Stores every test run with its examples, prompt version, model and pass marks.
- The Proof Console; the runs are published on our evidence page.
- Guardrails
- Keeps a tool from reaching anything but the approved model.
- Sealed mode, tested on every release; an audit log in your own database.
Every model platform, local runtime and speech service we use
AI models and model platforms
- Claude APIDefault
Model calls in our own products and in client tools
- Gemini API and Google AI StudioDefault
Model calls and prototyping; Stafflo's published test runs use a Gemini model
Test runs - OpenAI APIAlso
Model calls where a client already contracts with OpenAI
- Amazon Bedrock and Bedrock AgentCoreDefault
Models and agents inside the client's AWS account
- Gemini Enterprise Agent Platform (formerly Vertex AI)Default
Models and agents inside the client's Google Cloud
- Microsoft Foundry (Foundry Models, including Azure OpenAI)Also
Models and agents inside the client's Azure
- Hugging FaceDefault
Open-weight models, their licences, and hosted inference
- Open-weight modelsDefault
Models that run on the client's own machines, chosen per task and licence
Local and on-device AI
- OllamaDefault
Running open-weight models on a laptop or office machine
- LM StudioAlso
Trying and comparing local models on a workstation
- llama.cppAlso
Lean serving on CPUs, Apple silicon and small GPUs
- vLLMAlso
Serving open models to many people at once on GPUs
- NVIDIA NIM and TensorRT-LLMIn your stack
High-throughput serving on NVIDIA servers
- Apple Foundation Models frameworkAlso
On-device language features in iPhone apps
- OCR and document readingDefault
Scans, faxes and phone photos into fields, with the page and line kept
Speech and voice
- Speech-to-text on the phone or an office machineDefault
Telugu, Hindi, Tamil and other Indian languages, plus English, Spanish, French, German and Arabic, with no audio leaving your network
- Open-weight speech models from AI4Bharat, OpenAI and MetaDefault
Transcription and voices that run on your own machines or phones, chosen per language and checked for a commercial licence
- ElevenLabs speech-to-text and text-to-speechAlso
Transcripts and natural voices in Indian and world languages, for work whose audio is allowed to leave your network
- Gemini speech-to-text and text-to-speechAlso
Transcripts and generated voices through the Gemini API, or inside your own Google Cloud project
- Sarvam speech-to-text and text-to-speechAlso
Indian-language transcripts and voices from an Indian provider, or self-hosted in your own AWS account
- Amazon Transcribe, Azure Speech in Foundry Tools and Google Cloud Speech-to-TextAlso
Speech services inside your own cloud account; Amazon Transcribe in Mumbai and Azure Speech in Central India keep audio in India
Every test run on this site that uses a model names it. See the test runs
Parts your IT team already knows
Each tool ships as containers holding its engines, with Postgres for records and the audit log, Redis for queues, and the model you approve. We deploy it with Terraform into your cloud account or with Docker Compose onto your machine, behind your company sign-in, with logs going to your own monitoring.
Services
TypeScript on Node.js, Hono for APIs, Go for heavier services, Python for statistics.
Screens
Next.js and React, in your brand when a module sits inside your app.
Records
PostgreSQL, with search by meaning in the same database; Redis for queues.
Packaging
Containers; a REST API and a small JavaScript SDK for modules that sit in your own apps.
Deploying
Terraform into your cloud account; Docker Compose onto one machine; Kubernetes when you already run it.
Sign-in
Your company sign-in: Microsoft Entra ID, Okta or Google Workspace.
Monitoring
Logs and traces into your own tools.
Written in VS Code and Claude Code, and reviewed in pull requests on GitHub.
We build field apps for Android in Flutter, with speech transcribed on the phone.
Every part we build with
Default is what we reach for first. Also is what we use when the job calls for it. In your stack is what we work inside when your company already runs it.
Where we write and ship code
- Visual Studio CodeDefault
The editor for every repo
- Claude CodeDefault
Writing, refactoring and reviewing code in the terminal, IDE and browser; every change reviewed by the engineer who owns it
- ClaudeDefault
Research, writing, planning and design work
- GitHubDefault
Repos in the client's name from day one; pull requests and review
- GitHub ActionsDefault
Tests and checks on every pull request
- GitLab CI/CD, Azure Pipelines, Jenkins and Bitbucket PipelinesIn your stack
The same checks in your company's own pipeline, when it already runs one
- pnpmDefault
Workspaces and installs in our monorepos
- Docker containersDefault
Packaging each tool so it runs the same on a laptop, a server or a cloud
Web front end
- Next.jsDefault
Default for sites, apps and dashboards, including both studio sites
- TypeScriptDefault
Every front end and most services, typed end to end
- ReactDefault
Apps where Next.js's conventions or hosting don't fit
- HTML, CSS and JavaScriptAlso
Small, fast micro-apps with no framework
- Laravel with InertiaIn your stack
Front ends inside existing Laravel apps
- RemotionAlso
Videos rendered from code when a film must live in the repo
Mobile
- Swift and SwiftUIDefault
Native iPhone and iPad apps for the current iOS release
- App Intents, WidgetKit and Live ActivitiesDefault
Siri and Shortcuts actions, Home and Lock Screen widgets, live updates
- Flutter and DartDefault
One codebase for iOS and Android; our default for Android
- Kotlin and Jetpack ComposeAlso
Android-only apps that need deep device features
- App Store Connect, TestFlight and Google Play ConsoleDefault
Betas, releases and store review
Services and APIs
- GoDefault
Most back ends: APIs, workers and sync jobs
- HonoDefault
TypeScript APIs on Cloudflare Workers, Node.js or AWS Lambda
ig-comment-dm on GitHub - Node.jsDefault
TypeScript services, including the Stafflo engines
- REST with OpenAPIDefault
Documented APIs with generated, typed clients
- RedisDefault
Queues, locks and pacing for background work
- PythonAlso
Data work, statistical models and machine learning
- FastAPIAlso
Python APIs around a model or a data job
- DjangoAlso
Admin-heavy apps and existing Django systems
- Laravel (PHP)In your stack
Existing PHP systems: maintain, extend or move
- JavaIn your stack
Existing Java systems: maintain, or port to a current stack
Data and storage
- PostgreSQLDefault
The default database; audit logs sit in the client's own Postgres
- Neon (Lakebase Postgres, by Databricks)Default
Serverless Postgres with branching; both studio sites keep their own order data on it
Trust - pgvectorDefault
Search by meaning inside the same Postgres, with no separate vector store
- Amazon S3 and Cloudflare R2Default
Files, documents and exports
- Amazon Aurora DSQLAlso
Always-on, Postgres-compatible database on AWS, available in the Mumbai region
- Google BigQueryAlso
Analytics over large event and document data
- Amazon RDS and Aurora PostgreSQL, Cloud SQL and AlloyDB, Azure Database for PostgreSQLIn your stack
Managed Postgres in the client's own cloud account
- SQL Server, MySQL and Oracle DatabaseIn your stack
Reading from existing systems, and migrating off them
Sign-in, identity and payments
- devsforfun IDDefault
Our identity and billing platform: OAuth 2.1 and OIDC sign-in, Supabase-compatible tokens, checkout on both sites
- Sign in with GoogleDefault
Sign-in for consumers and Google Workspace users
- Amazon CognitoAlso
Users and sign-in inside the client's AWS account
- Microsoft Entra ID, Okta and Google Workspace sign-in (SAML, OIDC)In your stack
Company sign-in for every tool we deploy
- RazorpayDefault
Payments, with devsforfun ID as merchant of record
Infrastructure, delivery and monitoring
- TerraformDefault
Every cloud resource as code in the client's repo; nothing built by hand in a console
- Kubernetes (Amazon EKS, Google GKE, Azure AKS)In your stack
Deploying into clusters a client already runs
- Docker ComposeDefault
Single-machine installs on office servers
- OpenTelemetryDefault
Logs, traces and metrics sent to the client's own monitoring
- AWS Secrets Manager, Google Secret Manager, Azure Key VaultDefault
Keys and credentials kept out of code
- OpenTofuIn your stack
The open-source alternative to Terraform, when a client standardises on it
- Sealed modeDefault
Refuses every outbound connection except to the approved model; tested on every release
Trust
Analytics and machine learning
- pandas, NumPy and scikit-learnAlso
Data cleaning, analysis and small statistical models
- Jupyter notebooksAlso
Analyses a client's team can re-run
- Fine-tuning (LoRA on open models, or the provider's own tuning)Also
Only when tests show a prompt isn't enough
The five engines our tools are built from
Reconciler
Moves records between systems without losing or duplicating any, paced to each system's limits, holding the rest for replay.
- Accounted for (written, or held for replay)
- 100%
- Written twice
- 0%
- Run on 1 October 2026
- 10 examples
Extractor
Reads documents into fields and leaves a field empty rather than guess.
- Fields read correctly
- 99.16%
- Run on 1 October 2026
- 40 examples
Gatekeeper
Drafts replies and decides which ones a person must see first.
- Sent without review when they shouldn't have been
- 0%
- Run on 1 October 2026
- 100 examples
Consent Spine
Records who agreed to what, and deletes everywhere when they withdraw.
- Deletions completed in every system
- 100%
- Run on 1 October 2026
- 20 examples
Proof Console
Keeps the test results behind every number on the site.
It has no test run of its own: it keeps everyone else's.
Four engines have their own test script, run against Postgres and Redis; the Proof Console stores the results. See every test run
It works inside the systems you already pay for
Official API first. The vendor's own MCP server where it's generally available. A paced browser only where there's no API, and every action logged.
Microsoft 365 (Outlook, Teams, SharePoint, Excel)
DefaultMail, files and sheets through Microsoft Graph and the Work IQ MCP server
Google Workspace (Gmail, Drive, Sheets, Calendar)
DefaultMail, files and sheets through Google's APIs; Google's own MCP servers are in developer preview
Salesforce (Salesforce Platform, Agentforce)
AlsoApps on the platform, Agentforce agents, Hosted MCP Servers, and outside front ends and back ends that talk to Salesforce
HubSpot
AlsoCRM data through its APIs and the remote HubSpot MCP server
WhatsApp Business Platform
AlsoStatus updates and replies where customers already talk
Zoho (CRM, Books, Recruit)
In your stackSystems of record common in Indian firms
TallyPrime
In your stackAccounting data for Indian businesses
Bullhorn, Ceipal and JobDiva
In your stackCandidates, submissions and RTRs at staffing firms
HL7 v2, FHIR and X12 (835, 837)
In your stackCharts, claims and remittances for coding and billing firms
Freight software and carrier APIs
In your stackShipment records and milestones for forwarders and 3PLs
Pharma quality and document systems (such as Veeva Vault)
In your stackControlled SOPs and deviation records
DocuSign and Adobe Acrobat Sign
In your stackSigned RTRs, consents and agreements
Older systems too: Laravel and PHP, Java, and SQL Server, MySQL or Oracle databases, read where they are and moved only when it pays.
Tools that write into your systems do it through the Reconciler. It slows to each system's limit and holds what it can't write for replay; in its latest test run, no order was written twice.
Agent frameworks, automation tools and older systems
Agents, MCP and automation
- Model Context Protocol (MCP)Default
Connecting AI tools to a client's systems through standard servers
- Custom MCP serversDefault
Exposing a client's own system to their AI tools, behind their sign-in
- PlaywrightDefault
Browser automation for portals with no API, paced and logged; Roast My Website's page captures
- Claude Agent SDKAlso
Agents that read, write and use tools under fixed rules
- OpenAI Agents SDKIn your stack
Agents for clients on OpenAI
- Microsoft Agent Framework and Copilot StudioIn your stack
Agents for clients standardised on Microsoft
- n8n, Make, Zapier and Power AutomateIn your stack
Simple app-to-app automations, often the right first step
Legacy and migration
- Cloud set up by hand in a consoleIn your stack
Rebuilt as Terraform so it can be reviewed and repeated
- AI-built prototypes (Lovable and similar)Also
Turning a prototype into code a team owns and can run
- Older PHP and Java systemsIn your stack
Keeping them running while parts move to a current stack
What it connects to in your industry
IT staffing
- Connects to
- Outlook or Gmail; Bullhorn, Ceipal or JobDiva; vendor portals and VMS, by paced browser automation where there's no API; Excel and Sheets
- Runs on
- On a desktop in your office or in a cloud account in your name. Candidate data stays with you.
- Rules
- DPDP Act; New York City's rules on automated hiring tools; EU AI Act for EU clients
Medical coding
- Connects to
- EHR exports, HL7 v2 and FHIR, X12 835 and 837, clearinghouse and payer portals, scanned PDFs and faxes
- Runs on
- On a GPU workstation or a partner-supplied server inside your network. We measure it on your hardware first.
- Rules
- HIPAA through BAAs; AMA licensing of CPT; your clients' accuracy targets
Software agencies
- Connects to
- Your client's stack (Laravel with Inertia, Next.js, React, Vue, Django, Java); GitHub, GitLab or Bitbucket; Jira; questionnaire files
- Runs on
- In your client's cloud or on their servers, packaged as containers with a REST API and a small JavaScript SDK. Laravel and Next.js examples included.
- Rules
- Whatever your clients pass down: RBI outsourcing directions, insurers' rules, HIPAA
Lending
- Connects to
- Lender portals and APIs, paced; Android phones in the field; call recordings; WhatsApp; Excel
- Runs on
- Speech is transcribed on the phone or on a computer in your office. Recordings never go to an outside AI service.
- Rules
- RBI Digital Lending Directions 2025; DPDP, with consent recorded before capture
Financial services
- Connects to
- Loan origination and management systems, KYC records through your licensed providers, document stores, RBI directions and your SOPs
- Runs on
- In your data centre, or in an India-hosted cloud account in your name with the data kept in India.
- Rules
- RBI Digital Lending Directions 2025, FREE-AI, CERT-In six-hour reporting, DPDP, DORA for EU clients
Education
- Connects to
- LMS, student records, CRM, marketing lists, WhatsApp
- Runs on
- In your own cloud account or on a server on campus.
- Rules
- DPDP verifiable parental consent; no tracking or targeted ads aimed at children
Logistics
- Connects to
- Freight software, carrier APIs and websites (paced), bills of lading, invoices, packing lists, spreadsheets
- Runs on
- On an office server or in a cloud account in your name. Carrier connections run from your side.
- Rules
- DPDP for consignee data; declaration penalties
Pharma
- Connects to
- Quality and document systems (such as Veeva Vault), SOP libraries, deviation records
- Runs on
- On a server inside the plant network, with no outbound connection.
- Rules
- GMP, 21 CFR Part 11, EU GMP Annex 11, with your validation partner
From a demo to your environment
Most work starts with the audit, from $3,500, which picks the one process worth building first. Each stage adds only what its gate asks for.
- 1DemoA 15-minute callSynthetic data on a laptop, a hosted modelGate to the next stageOne process chosen
- 2PilotThe pilot, from $15,000One process in your environment in 4 to 6 weeks: your accounts, your sign-in, the audit log in your database, the source in your repoGate to the next stageThe pilot passes your tests
- 3In productionRun and improve, from $2,000 a monthEvery model update gated by your test set; a monthly note on what changedGate to the next stageA reviewer or regulator asks
- 4Under reviewPrivate AI readiness check, from $5,000And the security packLocal models on your machines in sealed mode; penetration tests commissioned on request and quoted separately
Check what you already pay for
Some work needs a setting in a tool you own, not a build. We'll tell you which, and here's where we'd look first.
Drafting, summaries and analysis for staff
- Try this first
- Claude Team or Enterprise, ChatGPT Business or Enterprise, Gemini in Google Workspace, or Microsoft 365 Copilot, whichever your data agreements cover
- When we'd build instead
- The work must follow your rules, touch your systems or leave an audit trail
Questions over your own documents
- Try this first
- A project or notebook in the assistant you already pay for
- When we'd build instead
- Answers must cite the version in force, and refuse when nothing covers the question
Moving data between apps
- Try this first
- n8n, Make, Zapier or Power Automate
- When we'd build instead
- Volumes hit rate limits, duplicates cost money, or the data can't leave your network
A first internal tool
- Try this first
- Lovable, or a spreadsheet with a form
- When we'd build instead
- It needs your company sign-in, an audit log, or to run in your cloud
CRM automation
- Try this first
- Agentforce in Salesforce, or HubSpot's built-in AI
- When we'd build instead
- The work spans systems outside the CRM
"Is our data safe in this AI plan?"
- Try this first
- Stafflo's free AI data privacy risk checker
- When we'd build instead
- It never needs a build. The checker shows what each AI plan does with your data; your own review makes the call
Interface and brand design is our sister studio's job: Pixelform.
Purchases on this site run on devsforfun ID, our own identity and billing platform.
devsforfun Studio is part of AWS Activate and Claude for Startups.
Reviewed 18 October 2026
Bring one process and one example
In 15 minutes we'll look at one process with you and show you the closest demo. The full picture, with the money worked out, is the audit.
Book a 15-minute call