Backend & platform engineer
I build systems that stay calm when thousands of people hit them at once.
I work on government digital services - taking public services live on WhatsApp at high concurrency, fine-tuning speech AI for Odia, and running my own Kubernetes platform on the side.
Selected work
Three systems I'm proud of.
Mini-Vercel
Self-hosted Kubernetes PaaS. Push a GitHub URL, get a live subdomain. Nixpacks + Kaniko builds, KEDA scale-to-zero, Nginx ingress.
WhatsApp Flow Generator
RAG-backed LLM tool that turns specs into valid WhatsApp Flow JSON. FastAPI microservices, PostgreSQL + Qdrant, Next.js chat UI.
Odia Whisper
Fine-tuned OpenAI Whisper for Odia speech recognition. Distributed inference backend with GPU autoscaling.
Experience
Bipros
Bhubaneswar, Odisha
Software Engineer
Dec 2025 - Present
Backend, DevOps
- Built langgenie.ai, a fault-tolerant, message-driven backend using RabbitMQ, PostgreSQL, and FastAPI with streaming results and job lifecycle management for a distributed Speech AI platform.
- Engineered a data pipeline to fine-tune OpenAI Whisper models for Odia language detection, transcription, and translation, achieving about 90% accuracy on regional datasets.
- Implemented priority queues and handled concurrent speech jobs with 2 GPU workers that auto-scale based on load.
- Built a RAG-based internal agent that converts PDF/image/text specs into production WhatsApp Flow JSON (bilingual English/Odia), cutting manual flow authoring for government services by 90%.
Bipros
Bhubaneswar, Odisha
Software Engineer Trainee
May-Dec 2025
Backend, DevOps
- Took 23+ live government services to a high-concurrency WhatsApp backend. Built distributed speech-AI infrastructure with GPU autoscaling, and shipped the platform's RAG document-to-Flow tooling.
- Engineered a backend processing 1,000+ citizen requests daily with zero data loss; cut processing time ~60% by replacing legacy portals with WhatsApp Flows across 150+ services.
- Fine-tuned OpenAI Whisper for Odia - detection, transcription and translation at ~90% accuracy - served through priority queues across auto-scaling GPU workers.
About
I'm a backend and DevOps engineer in Bhubaneswar, building the systems behind public digital services - speech AI, high-concurrency WhatsApp pipelines, and the infrastructure that runs them.
I care about software that stays quiet under load: fault-tolerant queues, clean job lifecycles, and platforms that scale to zero when no one's looking. Lately I've been fine-tuning speech models for Odia and running a self-hosted Kubernetes PaaS for the fun of it.
Stack