Flute of the Soul
Apurv Chudasama · AI Engineer
FAVORITES
Desktop About Projects Skills Journey Certificates Publications Resume Contact Blogs
Flute of the Soul
Desktop About Projects Skills Journey Certificates Publications Resume Contact Blogs
Flute of the Soul — Apurv Chudasama

AI Engineer · Ahmedabad

I build algorithms that listen.

Real-time voice agents, agentic architecture, and the unglamorous engineering that keeps them steady in production — sub-second latency, recoverable state, and calls that do not drop.

Every idea arrives as a new dataset, every setback as noise to be filtered. A bansuri is hollow: it sounds only when breath moves through the emptiness — which is the truest description of a voice pipeline I have found.

See the work Download résumé Get in touch
<700ms
End-to-end voice response, STT → LLM → TTS
1K+
Live calls handled in production
99.5%
Telephony uptime across Twilio and Telnyx
200+
Engineers taught across 5 colleges

At a glance

The facts, before the story

Current
Junior AI Engineer at Codiste and Dialora AI, Ahmedabad
Experience
2+ years in applied ML — production AI since June 2025, internships from 2023
Education
B.Tech Computer Science, CHARUSAT — CGPA 9.64, graduating December 2026
Based in
Ahmedabad, Gujarat, India — open to remote
Open to
AI/ML engineering roles and voice AI consulting
Recognition
Winner, DoseHack AI Hackathon (₹25,000) · Top 12, Codesahastra 11 · taught 200+ engineers

Works in

  • Python
  • C++
  • PyTorch
  • TensorFlow
  • LangGraph
  • LangChain
  • RAG
  • Transformers
  • OpenCV
  • YOLO
  • scikit-learn
  • XGBoost
  • FastAPI
  • PostgreSQL
  • Docker
  • Twilio
  • WebSockets
  • Jetson Nano

Selected work

Three systems, and what they had to survive

Each of these shipped and met a number. The interesting part was never the model.

Codiste · Voice AI

Real-time telephony voice agent

Twilio Telnyx WebSockets STT · LLM · TTS
Problem
Callers hang up when a bot pauses. Anything above a second of silence reads as a broken line.
Approach
Continuous voice streaming over WebSockets with barge-in handling, so the model can be interrupted mid-sentence the way a person can.
Impact
Sub-700ms end-to-end across 1K+ calls at 99.5% uptime.

Codiste · Agentic systems

Production LangGraph agent with durable memory

LangGraph RAG PostgreSQL FastAPI
Problem
Agents that lose their state on restart cannot be trusted with anything a customer paid for.
Approach
Persistent checkpointed state and conversational memory in PostgreSQL, with dynamic context retrieval instead of a growing prompt.
Impact
Conversations resume across restarts; retrieval replaced unbounded prompt growth.

NullClass · Computer vision

Real-time video models, shipped as apps

PyTorch OpenCV Streamlit
Problem
Age detection, sign-language prediction, and traffic analysis are demos until someone can actually use them.
Approach
Real-time video processing behind interactive Streamlit front-ends, deployed rather than notebook-bound.
Impact
Three deployed models non-specialists could operate directly.

How I work

Every build is a training loop

Ideas arrive as datasets, problems as noise to filter. The work is running the loop until it converges — then shipping it.

  1. 01

    Collect

    Understand the real constraint before writing anything. Latency budget, call volume, what breaks at 2am.

  2. 02

    Filter

    Strip the noise. Most "AI problems" turn out to be retrieval, state, or plumbing problems wearing a costume.

  3. 03

    Iterate

    Ship small, measure honestly, and let the numbers argue. Every interaction refines the next version.

  4. 04

    Converge

    Done is when it holds under real traffic — not when the demo works once on a good connection.

Writing

Notes from the build

Loading recent posts…

Everything here

Nine windows

Open any of them — they run as windows on this desktop.

Flute of the Soul

Building something that has to listen?

I take on voice AI and agentic systems work — the kind that has to hold up under real traffic. Ahmedabad, and remote.

apurvchudasama.edu@gmail.com Résumé

© 2026 Apurv Chudasama

Every ending is just a new model initialization.