SUCCESS STORIES

Systems we built, and what they changed.

Real problems, explained in plain language: AI, cloud infrastructure, real-time systems, search, digitization and automation. For each one we explain what the problem was, what we built, and what changed once it was live.

AI · RAG · On-premise New

A private AI assistant that runs with no GPU and no internet

We gave a government agriculture portal an AI assistant that answers farmers' questions from official scheme documents. It runs entirely on the portal's own locked-down server.

Farmers needed clear answers about government schemes: who qualifies, what share they pay, which documents to submit. Those answers were buried in long PDFs full of tables and abbreviations. The usual AI toolkit wasn't an option, because the server had no GPU, no Python, no Docker and no internet access after setup.

We built the whole assistant in Node.js. A compact open-source language model runs locally on the CPU. Search combines meaning-based matching with exact keyword matching, so scheme acronyms aren't missed. A reranking step then picks the most relevant passages, and strict prompting keeps the model answering only from the documents. All of it fits inside the existing app, starts automatically on reboot, and uses a fraction of the server's memory.

Node.js · llama.cpp · Qwen2.5-3B · Qdrant · BGE embeddings & reranker · MiniSearch (BM25) · Angular

Read the engineering write-up ↗ by Sidhant Tiwari

What it delivered

100% offline
No data leaves the server, and no paid AI APIs
0 GPUs
Runs on an ordinary 8-core CPU server
~4.3 GB
Total RAM for the whole AI stack, out of 16 GB
Cloud · AWS · Linux Networking New

30 users, one server, 30 separate network identities

We made a single AWS server send each user's traffic from that user's own pre-registered IP address. That replaced what would otherwise have been a fleet of 30 servers.

The platform connected to an external service that checks the source IP of every request against an address registered for each client. A mismatch means rejection. Giving every user their own server would have multiplied cost and maintenance. The obvious shortcut of binding each request to an IP in code looks right but silently fails on a server with several network interfaces.

We set up one instance with multiple network interfaces and a mix of IPv4 and IPv6 addresses to keep costs down. Each address got its own routing rules, so its traffic leaves through the right interface. We then tracked down three failures that only appear in production: silently rejected routes, a periodic network announcement that scrambled IPv6 traffic every half hour, and an overnight system update that dropped addresses. Finally we added a self-healing check that restores the setup within minutes if anything drifts, plus a test that proves every user's traffic leaves from the right address.

AWS EC2 · ENIs · IPv4/IPv6 · Linux policy routing · netplan · systemd · Node.js

Read the engineering write-up ↗ by Abhishek Tiwari

What it delivered

30 → 1
One server instead of one per user
0 leaks
All 41 addresses verified to exit from the correct IP
5-min
Self-healing checks that restore routing automatically
Custom Software · Real-time Systems

A trading platform that connects many brokers through one screen

We built an algorithmic trading system that watches live market data and acts on it in milliseconds — across multiple brokers at once.

Traders usually juggle separate apps for each broker, each with its own quirks and delays. We designed a single platform that plugs into all of them behind one clean interface, streams live market data without lag, and stays up even under heavy market load. The hard part wasn't the screens — it was the plumbing underneath: keeping data flowing fast and reliably when every millisecond counts.

Node.js · Angular · Redis · WebSockets · real-time data streaming

What it delivered

Multi-broker
One interface for several brokers at once
Low-latency
Live market data with minimal delay
Always-on
Built to stay up under market load
Digitization & Workflow

Taking a paper-heavy institution fully digital

We mapped how work actually flowed through a large institution and rebuilt it as software — cutting out millions of sheets of paper a year.

Big organizations often run on paper forms, physical approvals, and manual registers. We studied the real workflow end to end, then built digital modules that replaced each paper step — submissions, approvals, record-keeping — inside one connected system. The result wasn't just less paper; it was faster turnaround, a clear audit trail, and work that could happen from anywhere.

Node.js · Angular · Python · MongoDB · Redis · workflow automation

What it delivered

4M+
Sheets of paper saved per year
Paperless
End-to-end digital workflow
Auditable
Every step tracked and traceable
Legacy Modernization

Moving an old database to a modern stack — without losing anything

We migrated a large system from an old-style database to a modern one, cutting the rebuild time by two-thirds and keeping every record intact.

Aging systems get slow, costly, and risky to change, but the data inside them is too valuable to lose. We re-architected the backend, redesigned the data model, and scripted a careful migration from a traditional relational database to a modern NoSQL setup — then layered in caching and smarter loading so the new system felt fast for everyday users. Nothing was lost in the move, and the rebuild took a fraction of the expected time.

Node.js · Python · PostgreSQL → MongoDB · Redis · caching

What it delivered

−66%
Reduction in rebuild time
Zero loss
Every record migrated safely
Faster
Snappier for everyday use
Automation & Reporting

An accounting suite that does the busywork for you

We built a complete accounting and payroll system with live dashboards and automatic document generation — replacing hours of manual work.

Finance teams lose time to repetitive tasks: filling registers, calculating payroll, generating the same documents over and over. We designed an end-to-end package — expense and allotment registers, loan tracking, and payroll — with a live dashboard that shows where the money is going, and automatic bulk document generation that produces what used to be typed by hand. The backend, database, and reporting were designed as one system, so everything stays in sync.

Django · PostgreSQL · ApexCharts · automated document generation

What it delivered

Full suite
Expenses, loans & payroll in one place
Live
Dashboards for spend & allocation
Automated
Bulk documents generated instantly

Your project could be the next one here.

Freelance or contract, in India or abroad: if you have a problem that needs real engineering, let's talk.

info@zarvatech.in