ZenMagix Pvt Ltd
AI engineering & automation
Mumbai, India
Mumbai — IST
Agents that take real actions. Retrieval that cites its sources. Automation that does not fail silently at 3am. We are an engineering team, not a prompt shop.
We came from blockchain infrastructure — wallets, exchanges, smart contracts. Domains where a bug does not produce a bad user experience, it produces a loss. That is the instinct we brought with us: idempotency, auditability, least privilege, and the assumption that someone will try to break it.
An AI agent with tool access is a distributed system with a non-deterministic planner attached. We have built the first part for years.
AI-powered web applications: streaming interfaces, in-app retrieval, cost control and evals in CI.
Mobile and cross-platform apps with AI at the core, not bolted on afterwards.
Line-of-business systems built to your process, owned outright by you.
Agents that take actions, with approval gates, run traces and idempotent retries.
Retrieval that cites its sources, measured with a real evaluation harness.
Voice agents that handle calls in Hindi, Marathi, Tamil and English.
The implementation layer above WhatsApp Business API, where the SaaS tools stop.
n8n and custom workflows engineered so they do not fail silently at 3am.
Engineering-led SEO: crawl budget, rendering, structured data, Core Web Vitals.
Getting cited by ChatGPT, Perplexity and AI Overviews — measured, not promised.
Paid acquisition run against contribution margin, not vanity ROAS.
The full growth stack, coordinated rather than sold in disconnected retainers.
Five clusters, from AI engineering to robotics. See the full map.
Deliberately narrow. Everything here rests on the same substrate — retrieval, evaluation, and systems that hold state correctly under load — which is why we can do all of it properly rather than some of it loudly.
How we work
We start with the business outcome, not the model. Often the honest answer is that a rule, a form or an off-the-shelf tool solves it for a fraction of the cost — and we would rather say so in week one.
A narrow proof of concept against your real data, with an evaluation harness and a measured baseline. Demos on clean inputs prove nothing; this is the step that tells you whether to continue.
Production engineering in tight increments. Idempotent actions, least-privilege tool scopes, run traces you can replay, and graceful behaviour when the model is wrong or the provider is down.
Models get deprecated, indexes go stale, providers change pricing. The accuracy you measured in March is not the accuracy you have in September. We keep measuring.
In practice
Named client case studies are being written up now — several need sign-off before we can publish figures. Ask on a call and we can walk you through specifics under NDA.
A scoping call costs you nothing and usually ends with a clearer problem statement, whether or not you work with us.
