AI agents and chatbots that know their limits.
AI agents, chatbots and LLM apps that answer from your own documents and systems, take on the busywork and hand decisions back to a person. We built EVA, the assistant inside Effred One, and bring the same rules to your project: permissions respected, personal data masked, nothing changed without a confirm.
Who it’s for, and what to expect.
For companies that want AI to take on narrow, repetitive work, such as answering questions from documents, reading invoices and forms, triaging tickets and drafting replies, without losing control of what it does. We are based in Gurugram and work with clients across India and abroad. We also built EVA, the assistant inside our own Effred One platform.
Most projects start with a short paid proof of concept on your real data. It ends with measured accuracy, a plan and a fixed estimate for the production build, including expected model usage costs. Production work then runs in two-week sprints, each ending with a demo. We compare models on your data before choosing one, can mask personal identifiers such as PAN and Aadhaar before anything reaches a model, and can run open-source models on your own servers when data must not leave. The rules we build on are in our guide to what an AI assistant should never do.
What's included
Pick the parts you need, and add more once the first one is working.
AI agents
Agents that look things up, fill forms and prepare actions in your systems — a person confirms before anything changes.
Chatbots & assistants
Customer and employee chatbots for your website, WhatsApp, Slack or Teams, connected to your business systems.
RAG over your documents
Answers from your policies, contracts and manuals, with sources shown so people can check them.
AI automation
Classify, extract, summarise and route documents, emails and tickets inside the workflows you already run.
Evals & guardrails
Test sets and accuracy tracking, prompt-injection defences, and PAN, Aadhaar and account numbers masked before a model sees them.
The right model for each job
Claude, GPT, Gemini, Mistral or open-source models on your own servers — compared on your data for quality and cost.
How we work
Discover
A short paid discovery: goals, constraints, a written plan and estimate.
Design
Architecture, UX and a clickable prototype you can react to.
Build
Two-week sprints, with a demo at the end of each one.
Run
Launch, monitor and hand over — or we keep running it 24×7.
Engineering standards
What you can hold us to on every engagement.
Tech stack
Anthropic Claude, OpenAI, Google Gemini, Mistral, Groq, OpenRouter · self-hosted vLLM · RAG over your documents · Python · PostgreSQL · Databricks · Slack, WhatsApp Business
- Proof of concept on your dataMeasured accuracy, a plan and a fixed estimate for production, including expected model usage costs.
- Evals before productionTest sets and accuracy tracking for every use case; models compared on your data for quality, speed and cost.
- Permissions and maskingThe AI sees only what the person asking may see. PAN, Aadhaar and account numbers can be masked before anything reaches a model, with prompt-injection defences in place.
- A person confirmsAgents prepare actions and nothing changes without a confirm. Documents and records are data, never instructions: the rules we built EVA on.
- Data that must not leaveOpen-source models self-hosted with vLLM on your own servers.
Questions about AI development
The best results come from narrow, repetitive work: answering questions from documents, reading invoices and forms, triaging tickets and drafting replies for a person to approve. We start by finding two or three use cases with clear value — and we tell you when AI is not the right tool.
It depends on the use case, the data and how many systems are involved. We usually start with a short paid proof of concept on your real data. It ends with measured accuracy, a plan and a fixed estimate for the production build, including expected model usage costs.
A proof of concept usually takes a few weeks. Production — integrations, permissions, evals and monitoring — depends on scope and is planned in two-week sprints, each ending with a demo.
We design for it from day one. The AI only sees what the person asking is allowed to see, personal identifiers such as PAN and Aadhaar can be masked before anything reaches a model, and open-source models can run on your own servers when data must not leave.
Whichever fits the task: Anthropic Claude, OpenAI, Google Gemini, Mistral and others, or open-source models self-hosted with vLLM. We compare them on your data for accuracy, speed and cost before choosing.
Yes. Agents that work inside your systems, and chatbots for your website, WhatsApp or internal tools — always with clear limits on what they may do on their own.
Guides on AI development
Let's talk about AI development.
Software, people or growth — a person replies within one working day.