PROJECTS

CALEPOFT

March 2026 – Present

As someone who lost over 20kg in the last two years, I can tell you that keeping track of my fitness, my activity and my diet directly helped me tremendously. Knowing my calories, macros and patterns of physical activity was what distinguished my efforts from actual results. I still do have a way to go, but being the natural foodie I am, if I don't weigh or track my food, I know I risk indulging far more than I should and hence I combined my love for full-stack engineering with this problem that I had which led to the creation of CalepoFT.

This application is production-grade full-stack fitness platform built on a microservices architecture with event-driven async communication, observability tooling, and a normalised relational database — live at calepoft.com.

I designed it for users who would like to know what are their set targets to achieve with their diet. Depending on a user's goals and physical attributes, a nutrition target is customized for them, with the TDEE calculator guiding the estimations. From there, they can log their consumption throughout the day and even create meals that they have on the regular, so that they can be logged quickly.

• Built the platform using Next.js 16 App Router with TypeScript, structuring authenticated routes via route groups, implementing server and client components, and using Next.js API route handlers as a proxy layer between the React frontend and backend microservices.
• Designed a normalised PostgreSQL schema with ACID-compliant transactional guarantees (BEGIN/COMMIT/ROLLBACK), leveraging complex JOIN-based aggregations and nested JSON responses to efficiently serve structured API workloads across services.
• Built an event-driven system using Apache Kafka for asynchronous communication, implementing idempotent event processing alongside schema validation (Zod) to ensure correctness under at-least-once delivery, retries, and message replays.
• Orchestrated a multi-service Docker Compose environment with Nginx reverse proxying, enforcing service isolation, health checks, and secure non-root container execution to improve deployment reliability and operational robustness.
• Implemented an observability stack using Prometheus and Grafana, instrumenting services with metrics and logs to support performance monitoring, debugging, and reliability analysis across the distributed system.

CONVERSA

May 2025 – July 2025

In plain terms: Conversa listens in on a live business call, transcribes it as it happens, and uses AI to suggest what to say next. Those suggestions surface through a chatbot interface, so whoever is on the call has timely, relevant prompts at hand without having to think of everything themselves.

It's a real-time business conversation assistant leveraging AWS serverless infrastructure and LLM-powered suggestion generation to deliver timely, contextual insights during live business calls via a chatbot interface.

• Integrated AWS Transcribe and Lex for speech-to-text processing, with Lambda and API Gateway powering serverless backend computations for low-latency transcription and intent detection.
• Orchestrated LLM calls on Groq Cloud using prompt templates to generate timely, actionable, and impactful suggestions in a business context from live transcript data.
• Trained a HuggingFace multi-class text classifier (Zero-Shot Classification) to label buffered transcript segments, driving the trigger logic for delivering contextual suggestions on the chatbot interface.