Voice cloning and STT/TTS: what I tried and what worked
Benchmarks, trade-offs, and why moving off Colab mattered.
Hi, I'm Harshit, a backend engineer in Delhi NCR working with Python, FastAPI, RAG pipelines and voice AI. I like owning a service end to end, from data model to deployment.

Researched voice cloning and benchmarked speech engines for a real-time conversational product.
The platform's AI avatar needed a natural, consistent voice and fast speech handling. Early prototypes on Google Colab had inference too slow for live conversation.
Inference became fast enough for the conversational pipeline, and the team had benchmark data to choose an STT/TTS stack on evidence.
An AI chatbot that answers customer queries with grounded, on-brand responses from a live knowledge base.
A brand needed support answers that stayed accurate as product information changed, without hand-maintaining an FAQ.
A low-latency API the frontend could plug into directly, with knowledge that refreshes automatically.
Cut a barcode and OCR scanning endpoint from about 24 seconds to about 8 on an AI food-analysis platform.
Users scanned a product and waited around 24 seconds for results, far too long for a mobile scanning flow.
Response time dropped to about 8 seconds, roughly 3× faster, alongside personalized health insights built from user profiles.
Candidate workflows, interview scheduling and LLM-powered resume summaries, containerized with Docker and automated email via SMTP.
Tenant-aware access control with data isolation and dynamic role-permission mapping across applications.
Real-time recommendations and itinerary planning, combining LLMs with live data from SERP APIs.
I'm a backend engineer based in Delhi NCR, with a B.Tech in Information Technology. I got into backend work because I like the parts of software nobody sees but everybody depends on: the data model, the API contract, the query that has to come back fast.
Currently I'm at Webmobril Technologies, building APIs for real-time features, search and scheduling. Alongside that, I've gone deep on applied AI, from RAG pipelines over vector databases to benchmarking speech engines for a voice assistant. The work I enjoy most is owning a service end to end, from schema to deployment.
Right now I'm writing up what I learned from voice AI R&D, sharpening my problem-solving fundamentals, and looking for a team that's building something real with LLMs.
Benchmarks, trade-offs, and why moving off Colab mattered.