Software Engineering Experience at Family Shell LLP

I worked as a Software Engineering Trainee Intern at Family Shell LLP, contributing to two distinct applications: FamilyShell, a family management platform, and VoiceConnect, an AI-powered calling agent platform.

My experience covered production backend debugging, feature improvements, API investigation, automated testing, conversational AI engineering, audio processing, and technical research. Working across these applications gave me experience with both an existing production system and the development of an AI-powered voice application.

Role: Software Engineering Trainee Intern

Duration:May 2026 – October 2026

1. FamilyShell — Family Management Platform

FamilyShell is a family management platform. I contributed to its existing production application, focusing on investigating and resolving backend bugs, improving feature behavior, and validating fixes across backend services and client interactions.

Production Engineering Contributions

• Investigated and resolved production backend bugs across APIs, database operations, authentication, reminders, notifications, and file storage workflows.
• Traced frontend-to-backend request flows by analyzing backend logs, Flutter logs, database records, and REST API responses to identify root causes.
• Implemented backend fixes involving the Health Records module, reminder scheduling, notification delivery, authentication workflows, email reminders, and attachment management.
• Worked with JWT authentication, Redis, Cloudflare R2 object storage, WebSocket notifications, and scheduled background jobs.
• Performed regression testing after bug fixes to verify expected behavior across backend APIs and Flutter clients.
• Collaborated with the development team using Git-based workflows to validate and deliver backend improvements.

Features and Areas of Work

• Health Records API
• File attachments and Cloudflare R2 storage
• Reminder scheduling
• Notification delivery
• Email reminder workflows
• JWT authentication
• Background scheduling
• API validation and error handling
• Regression testing

Technologies Used

Python, FastAPI, MongoDB, Redis, Docker, Cloudflare R2, JWT authentication, Flutter, REST APIs, WebSockets, Git, and GitHub.

Key Learning Outcomes

• Debugging and maintaining an existing production backend.
• Using application logs and API responses to investigate issues across system components.
• Validating backend fixes through regression testing.
• Understanding authentication, storage, notification, and scheduling workflows in a production application.
• Collaborating on a shared codebase through Git-based development workflows.

2. VoiceConnect — AI Calling Agent Platform

VoiceConnect is an AI-powered calling agent platform designed to help businesses handle customer calls. The agent uses company knowledge to answer relevant questions, maintains conversational context, and can raise support tickets when it cannot resolve a customer's query.

My contributions focused on improving the existing knowledge retrieval architecture, developing the conversational reasoning engine, building a provider-agnostic audio layer, improving speech turn detection, and researching architectural approaches for retrieval and telephony integration.

Provider-Agnostic Vector Database Layer

• Refactored the existing Pinecone-specific vector database integration into a provider-agnostic vector store layer.
• Implemented configurable vector store selection and client creation to reduce the system's dependence on a single vector database provider.
• Refactored retrieval and document ingestion components to use configurable embedding and vector store integrations.
• Centralized configuration and validation for language models, embedding providers, and vector databases.
• Added strict configuration validation to detect conflicting provider definitions and prevent silent defaults.
• Implemented ephemeral source-document processing so uploaded files are deleted after ingestion rather than retained by the platform.

Conversational Agent and Tool Calling

• Developed a provider-agnostic conversational reasoning engine supporting Gemini, OpenAI, and Claude integrations.
• Implemented sliding-window conversation memory that retains complete human-and-AI conversation turns.
• Implemented multi-round tool execution with error handling and configurable execution limits.
• Separated the platform's telephonic behavior rules from client-specific runtime configuration.
• Developed a company knowledge search tool that connects the conversational agent to the existing RAG pipeline and provides source-attributed results.
• Implemented a support-ticket creation tool with contact validation and safeguards for tool execution.

Provider-Agnostic Audio Processing

• Implemented abstract interfaces for speech-to-text and text-to-speech providers.
• Integrated Deepgram Nova-3 for speech recognition and ElevenLabs Flash v2.5 and Murf Falcon 2.1 for speech synthesis.
• Developed a voice pipeline coordinating microphone input, speech recognition, the conversational agent, speech synthesis, and speaker playback.
• Implemented dynamic echo cancellation as part of the audio pipeline.
• Added tool-aware language-model response streaming and pipelined text-to-speech playback to support more responsive voice interactions.

Real-Time Speech Turn Detection

• Implemented dual-gate utterance completion using Deepgram's acoustic endpointing and UtteranceEnd events.
• Configured 700-millisecond acoustic endpointing as the primary completion mechanism and a configurable 1,000-millisecond word-gap fallback.
• Added buffer-aware handling of empty speech-final events to prevent buffered transcript text from being discarded when it should complete a turn.
• Reset trailing-event suppression flags during unmuting to prevent stale state from affecting subsequent caller utterances.
• Added safeguards against delayed events incorrectly flushing partial transcripts during multi-chunk utterances.
• Wrote regression tests for missing completion events, event handling, and turn-boundary scenarios.

Technical Research and Documentation

• Researched and documented 23 RAG strategies, evaluating their suitability for VoiceConnect's business knowledge retrieval and voice-based customer support requirements.
• Researched and documented an implementation approach for integrating Exotel as the telephony layer connecting telephone calls to the speech recognition, conversational agent, and speech synthesis pipeline.
• Documented the system architecture, configuration options, audio processing flow, speech turn detection, and automated testing approach.

Automated Testing

• Developed automated Pytest tests for the conversational agent, audio components, configuration, and speech turn detection.
• Used mocked external service clients to test application behavior without consuming live API credits.
• Added regression tests to verify fixes for configuration errors, conversation handling, and speech-event lifecycle issues.

Technologies and Tools

Python, LangChain, RAG, vector databases, embedding models, Gemini, OpenAI, Claude, Deepgram Nova-3, ElevenLabs Flash v2.5, Murf Falcon 2.1, Pytest, environment-based configuration, Git, and GitHub.

Overall Experience

Working on FamilyShell and VoiceConnect exposed me to two different areas of backend engineering: maintaining and improving an existing production application, and developing the core components of an AI-powered voice platform.

These experiences strengthened my skills in Python backend development, debugging, API integration, configuration design, automated testing, conversational AI, audio processing, and technical research.