Rohan Sharma

Rohan Sharma, voice AI engineer

I build the streaming systems behind voice agents, from the first audio frame to the first word of the reply.

Email meSee the work
Notes

Say something

Say anything, then see what your voice measures.

Range
-
Pace
-
Pauses
-
Ending
-

Audio stays in this tab. Nothing is recorded or sent.

The 200 ms gap

People pause about 200 milliseconds between turns. A voice agent has to hear you finish, decide and answer inside that, or it stops feeling like a conversation.

Gap between turns: Stivers et al., PNAS, 2009.

An 1859 phonautograph: a horn that focuses sound onto a membrane and a stylus that traces the sound wave onto a rotating cylinder.
Scott de Martinville’s phonautograph, 1859, the first machine to draw a sound wave. Public domain, via Wikimedia Commons.
A plucked violin string and a piano note as amplitude envelopes: the shape of loudness over time that a turn detector watches. By ItzaGarciaOrdonez, CC BY-SA 4.0, via Wikimedia Commons.

Work

Try OmniToken

Type anything. This is the Go library itself, compiled to WebAssembly.

o200k_base

Tokens

Where is the nearest pharmacy?

Token IDs

  • 11977
  • 382
  • 290
  • 35210
  • 43125
  • 30
Tokens
6
Characters
30
Encode time
-

Click the box to run it live. The first use downloads the tokenizer, about 2.4 MB.

Then see how fast it is natively.

ns/op and allocs/op: cl100k_base, JSON input. Speedup: geometric mean against tiktoken-go across the benchmark harness. Author’s harness on Windows 11, Intel i7-1250U, Go 1.24.2.
Operationns/opallocs/opSpeedup
CountTokens1,517015.84x
EncodeOrdinary1,661113.09x
Decode20422.29x

Use it in a Go service.

go get github.com/ron2111/omnitoken

Source on GitHub, the write-up on Medium.

About

Portrait of Rohan Sharma

I build voice agents for a living, and most of the work is the part nobody hears.

At Nugget by Zomato I own the voice agent platform. Live audio comes in, the agent decides what to do, and a reply goes out, all inside a latency budget measured in milliseconds. That means voice activity detection, noise handling, turn-taking and orchestration, and a lot of care about what happens when something fails mid-call.

Army telephone operators at a large switchboard, wearing headsets and connecting calls by hand.
Army telephone operators at the Victory switchboard, Potsdam, 1945. National Archives, public domain, via Wikimedia Commons.

Before that I was tech lead at GoodSpace AI, where the same ideas ran at a different scale: 1,000+ concurrent calls, 25,000+ interviews in 55 languages, and a Kafka pipeline underneath it all.

On the side I write Go libraries and MCP servers, usually to scratch a performance itch. OmniToken started that way.

If you are building something with voice, I would like to hear about it.

Rohan SharmaVoice AI Engineer

Army telephone operators at a large switchboard, wearing headsets and connecting calls by hand.
Army telephone operators at the Victory switchboard, Potsdam, 1945. National Archives, public domain, via Wikimedia Commons.

Background

Experience

Writing

Community

  • Microsoft Learn Student Ambassador2021 to 2024
  • Google Cloud Ready Facilitator1,350+ students mentored

Education

  • Delhi Technological UniversityB.Tech, production and computer engineering2020 to 2024

Certifications and awards

  • Microsoft Certified: Azure FundamentalsAZ-900
  • Google Cloud Skills BoostSkill and quest badges
  • Salesforce Trailhead HackathonFirst prize
  • Seize the Trail QuestWinner
  • Hacktoberfest2020

Building something with voice?Tell me what it is.

Opens your email app with your message ready to send to 2111rohan@gmail.com.

Email
2111rohan@gmail.com
LinkedIn
linkedin.com/in/rohan-sharma21
GitHub
github.com/ron2111
Medium
medium.com/@2111rohan