Software Engineer at Vani.coach (2024-11 – Present)
Backend engineer with end-to-end ownership of backend services, AI/LLM features, and cloud infrastructure.
- Designed and shipped LLM-powered feedback generation across three rubrics (Articulation, Thinking, Accuracy), utilizing Azure Speech Pronunciation assessment and LLM prompting — currently producing ∼5,000 feedback generations/month in production.
- Built and shipped two production real-time voice agents on Pipecat/Daily with Azure Speech (STT/TTS), LLM-driven dialogue, and S3 audio persistence: an onboarding agent that replaced human coach-led discovery calls for ∼90% of incoming learners, and a roleplay agent for communication practice with structured per-session logging.
- Reduced Azure infrastructure spend 73% ($3,000 → $800/month) by consolidating services onto a single VM, right-sizing the database, and removing high-cost feature paths.
- Personalized LLM-generated practice scenarios by introducing a per-user Memory context injected into prompts; post-launch interviews with 10 users confirmed improved relevance.
- Designed and shipped a Plans and Subscriptions system end-to-end (versioned REST APIs, admin endpoints) with referral attribution for a Xoxoday B2C partnership; payments and self-serve activation intentionally deferred pending B2C traction.
Software Engineer at MSG91 (2023-04 – 2024-05)
Responsible for full stack development of MSG91's help center website offering 'KnowledgeBase'.
- Optimized page load speeds by over 10x and enabled instant browsing through custom caching.
- Consistently maintained a 95+ SEO score.
- Implemented semantic search for better article discovery.
- Added functionality to track views and gather reader's feedback, enabling continuous feedback on content.
- Developed a Chat-Bot answering customer support queries from Knowledgebase leveraging LLMs.
