Sourabh Sharma
Résumé

Software Engineer · AI & backend systems

Sourabh
Sharma

I build the backend and AI services behind real-time products.

Async Python APIs, event-driven messaging and model-serving microservices — from vision and speech models to RAG chatbots, taken from prototype to production.

Get in touchDownload résumé
Based in
Jaipur, India
Currently
Associate Software Developer, Predusk

01 Selected work

Systems I’ve built and shipped.

Personal project · AI voice companion

Flare

Live · invite-only

A voice-first AI companion with a face. Hold a button, speak, and a stylised 3D character listens, thinks and answers out loud with lip-synced speech — all orchestrated at the edge on Cloudflare.

  • Edge pipeline on a Cloudflare Worker (Hono) chaining Whisper speech-to-text, Llama 3.1 and Kokoro text-to-speech, streaming audio back while metadata travels in response headers.
  • Real-time lip-sync computed entirely in the browser — Web Audio FFT analysis drives 67 ARKit blend shapes, with no server-side phoneme aligner.
  • Rolling context summarisation keeps long conversations cheap, and D1 (SQLite) stores users, threads and messages with cascading deletes.
  • Provisioned with Terraform and shipped through GitHub Actions; the API is covered by 21 automated Vitest tests and strict TypeScript across the monorepo.
  • Next.js 15
  • React Three Fiber
  • Hono
  • Cloudflare Workers
  • D1
  • R2
  • Clerk
  • Terraform
  • Vitest

Predusk · RAG chatbot & AI services

Workplaceslm

Production · Predusk Technology

A retrieval-augmented chatbot platform and its supporting AI services. I rebuilt the real-time layer, removed event-loop blocking and upgraded the research pipeline.

Data flow · Workplaceslm
  1. Chat client
  2. EMQX (MQTT)
  3. FastAPI service
  4. Celery workers
  5. LangGraph agents
  6. Results streamed back over MQTT
  • Replaced raw WebSockets with an EMQX MQTT broker and a dedicated FastAPI microservice using JWT authentication for stable, persistent connections.
  • Removed event-loop blocking by moving psycopg2 and the MinIO SDK to asyncpg and aioboto3.
  • Offloaded compute-heavy deep search to Celery and streamed results to users over MQTT, eliminating HTTP polling.
  • Upgraded the deep-research pipeline from Agno to a multi-agent LangGraph design, and built Kokoro text-to-speech and Faster-Whisper speech-to-text microservices for voice.
  • Python
  • FastAPI
  • MQTT / EMQX
  • Celery
  • LangGraph
  • asyncpg
  • MinIO
  • Kokoro
  • Faster-Whisper

Predusk · Data Annotation & Delivery Platform

DADP

Production · Predusk Technology

Real-time event streaming and computer-vision services for an annotation platform, built to cut the manual effort annotators spend on images and video.

Data flow · DADP
  1. Annotation UI
  2. FastAPI + Celery tasks
  3. SAM 2.1 / YOLO11x + BoT-SORT
  4. RabbitMQ + MQTT events
  5. Live progress to the annotator
  • Replaced unstable Server-Sent Events with a RabbitMQ and MQTT event-streaming microservice for reliable progress tracking of long-running Celery tasks.
  • Built and deployed a GPU-accelerated FastAPI service around SAM 2.1 (Hiera Large) that turns bounding-box or click prompts into polygon coordinates.
  • Built an automated video-tracking service using YOLO11x detection, YOLO11x-cls appearance embeddings and BoT-SORT to follow objects across frames.
  • Python
  • FastAPI
  • Celery
  • RabbitMQ
  • MQTT
  • SAM 2.1
  • YOLO11x
  • BoT-SORT

Predusk · Web automation & scraping platform

Oceanmind

Production · Predusk Technology

An automation engine that produces cruise quotations, replacing a manual workflow and structuring free-text customer requests with an LLM.

Data flow · Oceanmind
  1. Quote request
  2. Playwright browser contexts
  3. Gemini request parsing
  4. PostgreSQL + MinIO
  • Migrated storage from SQLite and JSON files to PostgreSQL and MinIO.
  • Moved scraping from Selenium to Playwright and replaced per-request browser instances with isolated browser contexts, sharply reducing RAM use and improving speed.
  • Integrated the Google Gemini API to parse unstructured special requests from customer comments into structured data.
  • Python
  • Playwright
  • PostgreSQL
  • MinIO
  • Gemini API

Personal project · Text-to-speech

Multilingual TTS Microservice

Dockerised service

A standalone text-to-speech service on the Kokoro model with selectable languages and multiple voices, plus a small Streamlit front end.

Data flow · Multilingual TTS Microservice
  1. Streamlit front end
  2. FastAPI service
  3. Kokoro TTS
  4. Audio response
  • Dynamic language selection and multi-voice generation behind a FastAPI interface.
  • Containerised with Docker so it runs identically on any machine.
  • Python
  • FastAPI
  • Streamlit
  • Kokoro
  • Docker

02 Experience

Where I’ve been doing the work.

  1. Nov 2025 – PresentCurrent

    Associate Software Developer

    Predusk Technology Pvt Ltd

    Jaipur, India

    Backend and AI engineer on the Workplaceslm chatbot platform and the DADP annotation platform.

    Workplaceslm — RAG chatbot & AI services

    • Moved real-time chat from raw WebSockets to an EMQX MQTT broker, with a dedicated FastAPI microservice and JWT authentication.
    • Eliminated event-loop blocking by migrating psycopg2 and the MinIO SDK to asyncpg and aioboto3.
    • Streamed Celery deep-search results to users over MQTT instead of HTTP polling.
    • Upgraded the deep-research pipeline from Agno to a multi-agent LangGraph architecture.
    • Built Kokoro TTS and Faster-Whisper STT microservices for voice features.

    DADP — Data Annotation & Delivery Platform

    • Replaced Server-Sent Events with a RabbitMQ and MQTT streaming service for reliable task progress.
    • Deployed a GPU-accelerated SAM 2.1 service that generates polygons from box or click prompts.
    • Built a YOLO11x and BoT-SORT video-tracking service to reduce manual annotation effort.

  2. Jan 2025 – Oct 2025

    Data Analyst Trainee, AI & ML Engineering

    Predusk Technology Pvt Ltd

    Jaipur, India

    Worked on the Oceanmind web-automation platform within the AI & ML engineering team.

    Oceanmind — web automation & scraping

    • Built an automation engine for cruise quotations that replaced manual work, and moved storage from SQLite/JSON to PostgreSQL and MinIO.
    • Migrated scraping from Selenium to Playwright, using isolated browser contexts instead of a full browser per request to cut RAM use and speed things up.
    • Integrated the Google Gemini API to structure free-text special requests from customers.

03 Skills

The tools I reach for.

Languages

  • Python
  • JavaScript
  • SQL

AI & machine learning

  • LangGraph
  • Gemini API
  • vLLM
  • RAG
  • YOLO11x
  • SAM 2.1
  • BoT-SORT
  • OCR
  • Kokoro TTS
  • Faster-Whisper

Backend

  • FastAPI
  • Celery
  • REST APIs
  • JWT

Messaging & real-time

  • MQTT (EMQX)
  • RabbitMQ
  • WebSockets

Data & storage

  • PostgreSQL
  • asyncpg
  • Redis
  • SQLite
  • MinIO
  • S3

Frontend

  • Next.js
  • Streamlit
  • HTML & CSS

DevOps & automation

  • Docker
  • Linux deployments
  • Git & GitHub
  • Playwright
  • Selenium

04 About

Engineering for systems that have to stay up.

I’m a software engineer at Predusk Technology in Jaipur, working across Python backends and applied AI. Day to day that means FastAPI microservices, real-time messaging, and putting vision, speech and language models into production.

I like replacing the fragile parts of a system — polling, blocking calls, a full browser per request — with architecture that holds up under load. Outside work I build and ship small products end to end, such as Flare, a voice companion that runs on Cloudflare’s edge.

Focus areas

  • Async Python and FastAPI services
  • Real-time messaging with MQTT and RabbitMQ
  • Vision, speech and LLM pipelines in production
  • Edge deployments on Cloudflare

Education

Bachelor of Computer Applications (BCA)

SSG Pareek College, University of Rajasthan

2022 – 2025 · Jaipur, India

05 Contact

Let’s talk about what you’re building.

Got a role, a project or a question about anything above? Email is the quickest way to reach me — I’m based in Jaipur and happy to talk across time zones.