projects

A unified operations platform for electricity distribution utilities. · Stratalis Tech

Electricity distribution companies in India typically operate meter communication, meter data management, billing, field operations and consumer services as separate systems from separate vendors. Discom360 brings these functions onto a single cloud-native platform, from smart-meter data acquisition through billing intelligence to the applications used by consumers and field staff.

• Designed and built the platform from the ground up as co-founder, and took it to production.
• Shipped two mobile applications on the App Store and Play Store. The consumer app covers consumption, billing, payments and grievances. The field app supports installation and service crews, and is built to work offline.
• Smart-meter communication and ingestion built to the DLMS/COSEM and IS 15959 standards.
• Built to scale from pilot deployments to large rollouts on the same platform.

Point of sale and operations for small retailers and restaurants, now FloPOS. · RetailQ

A multi-tenant point-of-sale platform for MSMEs. It began as a POS for shops across food service, electronics, salons and liquor retail, and grew into a full restaurant platform with a kitchen display, loyalty, inventory, recipes, table and queue management, staff management, reporting and multi-outlet support. At its peak more than 190 franchises across Punjab ran on it.

• Designed the backend so that each store's services ran in isolation behind their own subdomain.
• Plug-and-play thermal receipt printing from the browser over WebUSB.
• OCR and LLM based product categorisation that halved manual catalogue entry.

PHPApacheAngularJSNext.jsDockerWebUSBPythonMistral

Can the way you use your computer prove that you are you, and no one else?

A multimodal behavioural biometrics dataset and continuous authentication system, built from 445 GB of competitive gameplay data. Released as an arXiv paper, an open dataset on Hugging Face and the logging software used to collect it, with further papers in preparation.

• 79 sessions from 28 participants and 102.5 hours of active capture, across mouse, keyboard, network, video and hardware metadata.
• Provided the GPU machines the gameplay sessions ran on, and supervised the collection lab whenever I was available.
• Built the data pipeline and the backend server participants used to upload their sessions.
• Aggregated, cleaned and released the dataset on Hugging Face, and co-authored the paper.

PythonPyTorchData PipelinesBehavioural Biometrics

Academic feedback as a continuous loop, not an end-of-term formality.

Live across Thapar Institute and used by more than 10,000 students. Students give feedback after individual lectures, tutorials and labs rather than once at the end of a course. They can also report how far they feel each course learning outcome has been met. Responses stay anonymous, and faculty see only the aggregate.

• Backend team lead. Led and designed the backend architecture.
• Role-based access control and state management across more than seven user types, built with Flutter, REST APIs and a LAMP backend.
• Analytics pipelines on Apache Kafka and modular PHP microservices, feeding a Next.js dashboard that gives the institute's academic leadership real-time visibility.
• Session-wise feedback for lectures, tutorials and practicals, with teacher-authored questions alongside the standard set.
• CLO-based feedback, giving a continuous view of outcome attainment through the semester.
• MCQs, descriptive answers, slider ratings and comprehension-based questions.
• Restricted to valid Thapar IDs. Shipped on both the App Store and Play Store.

PHPKafkaFlutterNext.jsMySQLApache

Automated timetable generation for the institute.

Genetic-algorithm scheduling that balances hard and soft constraints across the university's more than 25 departments, with a Next.js front end and a LAMP back end. Faculty can edit their timetables interactively, instead of waiting on revised PDFs. Funded by a US$17,000 grant.

• Backend team lead. Led and designed the backend architecture.
• Designed the system as a set of microservices and built all of its APIs.
• Owned data handling end to end, covering cleaning, consolidation and presentation across departments.
• Interactive editing made the faculty feedback cycle three times faster than the old PDF process.
• Guided by Prof. Rajkumar Gupta and Prof. Tarunpreet Bhatia.

PHPMySQLApacheNext.jsGenetic AlgorithmsMicroservices

Reading industrial machine screens with computer vision. · Trident Group

Sole developer of an on-site system that monitors PLC-controlled machines through their screens, running on low-powered devices. It tracks the relevant screen regions, extracts control values with OCR, and checks them for anomalies against domain-specific statistical rules.

PythonOpenCVTesseractMediaPipeNumPypandas

Serving models efficiently on mismatched hardware.

My homelab is a deliberately mixed set of machines, spanning NVIDIA and Apple Silicon with very different memory bandwidths. I use it to work out what actually makes local inference fast. That covers quantization choices, CPU/GPU offload, KV cache behaviour, and splitting a single model across machines that were never meant to work together.

• Heterogeneous and distributed serving with llama.cpp, vLLM and MLX.
• Quantization and MoE inference trade-offs, especially prompt processing versus generation throughput.
• Fine-tuning small models with Unsloth for narrow tasks.

llama.cppvLLMMLXCUDAGGUFQuantizationUnsloth

A MOOC platform for a state power utility's training institute.

Deployed a Moodle-based learning platform for the Technical Training Institute of Punjab State Power Corporation Ltd. (PSPCL), Government of Punjab, on open-source infrastructure.

MoodleLinuxPHP

Self-hosted clipboard sync across all your devices.

Copy on one machine and paste on another, without handing your clipboard to someone else's cloud. A monorepo with a Dockerised server and a Flutter client. Still a work in progress.

FlutterNode.jsRedisDockerGitHub Actions

Where most of this actually runs.

Linux servers, containers, VLANs, firewalling, a GPU machine and a set of self-hosted services. It is the reason I know the tools on my stack list, since most of them were learned by breaking something late at night and fixing it.

LinuxDockerLXCProxmoxKubernetesOPNsenseCloudflare