Read Chapter 1
Purushottam Chaudhary
Purushottam Chaudhary Founder & AI Inference Engineer, QuickDial AI Founder, InfiniSync Author, The Inference Mechanic
LinkedIn
Founder · AI inference engineer · Published researcher · Author

Passionate about taking intelligence to the smallest device that can hold it.

Fifteen years delivering software for GE Healthcare, S&P Global, Bristol Myers Squibb and State Street; three years training machine-learning models; three companies founded. At QuickDial AI I build the speech and language models and the platform they run on, so a business can be answered by a machine it already owns.

Get in touch About QuickDial
0 yrssoftware, healthcare to finance
0companies founded
0continents worked across
0languages given a voice
Now · Founder & AI inference engineer

QuickDial: a business that answers its own phone.

Excellence in inference engineering. QuickDial owns the whole voice call path — speech in, understanding, speech out — and runs it on commodity compute. AgentBox answers a business on the phone and the web; Quick TTS streams voices on ordinary hardware.

583 mstypical reply on live traffic
3languages, all built in-house
<2¢per minute today
Oct 2026first hotel pilots
quickdial.aiDeck
Founder · AI inference engineer · Built the stack

A voice that picks up, understands and replies — on hardware the business already owns.

I built the speech recognition, the language model and the voices in-house, the platform they run on and the hospitality product above it, and I run the company: models, team and customers. Co-founder Ankur Kapoor and I fund it equally; he joins full time on the round.

  • What I builtStreaming speech recognition, a grounded language model for tool use, and text-to-speech in English, Hindi and Spanish — plus segmented KV-cache prefill that cut time-to-first-token 7.5×.
  • Current statusHotel and restaurant pilots are live; a pre-seed round is open. A demonstration is available at hotel.agentbox.quickdial.ai; the numbers are in the deck.
Where I have worked

Enterprise systems, platforms, models — then a company of my own.

Each role contributed to the next: the discipline of regulated software, the scale of market-data systems, the leadership of a platform business, the research practice of generative 3D. QuickDial draws on all of them.

QuickDial — a product of EasyDial

Now · Jul 2026 –
Founder & AI Inference Engineer

EasyDial was set up in 2025, where the voice loop first ran end to end on one commodity box; QuickDial is the product it became. Built the speech, language and voice models, the multi-tenant AgentBox platform and the hospitality product.

Where everything below is now being spent.
Voice AIInferenceEdgeHospitality

InfiniSync

Jan 2024 –
Founder

IT consultancy in deep tech, AI, blockchain and software. Delivered backend platforms, integrations and ML systems for clients in India and the US; the affiliation on the ICU-mortality paper.

Outcome: a services business that funds product development, and the research collaborations behind the publications.
Deep techAIBlockchainConsulting

brahmGAN

Feb – Jul 2023
Co-founder & CTO

3D model generation from images via NeRF and Gaussian splatting — a generative-3D startup, technical lead from research to product.

Outcome: hands-on experience training models rather than only integrating them.
NeRFGaussian splattingGenerative 3D

KNNX Corp

Aug 2021 – Feb 2023 · India
Assistant Vice President · Head of Technology

Led blockchain platform development: architecture, team and delivery of an enterprise distributed-ledger platform.

Outcome: building and leading an engineering organisation, and delivering a platform used by other companies.
BlockchainPlatformEngineering leadership

S&P Global

Jun 2020 – Aug 2021 · India
Lead Engineer

S&P Dow Jones Indices platform — the systems that compute and publish the world's most-followed market indices.

Outcome: systems in which an incorrect figure has public consequences — rigorous measurement as standard practice.
Financial dataIndicesScale

GE HealthCare

Apr 2018 – Jun 2020 · Bangalore, India
Technology Lead

R&D at the John F. Welch Technology Centre, GE's largest research site outside the US — healthcare software from prototype to release.

Outcome: regulated healthcare software delivered from prototype to release.
HealthcareR&DMedical software

IT consulting, United States

Sep 2010 – Mar 2018 · United States
Information Technology Consultant

Seven and a half years consulting to GE Healthcare, Bristol Myers Squibb and State Street Bank — healthcare, pharma and financial-services systems, from inception to production support.

Outcome: seven years of delivering and accounting for work directly to enterprise clients.
GE HealthcareBristol Myers SquibbState StreetEnterprise
Research & publications

Two papers, two fields.

Journal · 2026 · Peer-reviewed

Conditioned Latent State-Space Network for Intervention-Aware ICU Mortality Prediction Using Longitudinal Physiological Data

Nitin Jain, Purushottam Chaudhary, Gaurav Jindal, K. Deepthi Reddy, Bharti Sharma, Dipika Jain

International Journal of Computer Information Systems and Industrial Management Applications, Vol. 18, No. 22s, pp. 1358–1376. A state-space model over longitudinal ICU data that conditions on interventions, evaluated on 2,000 MIMIC-III patients: AUROC 0.791, calibration error 0.072, with counterfactual and uncertainty analysis for clinical decision support.

doi:10.70917/ijcisim-2026-5533
Inference systems · 2026 · Preprint

Composable KV-Cache Segments and the Decode Wall: Voice Agents on Commodity Compute and Edge Devices

Purushottam Chaudhary

The two results QuickDial runs on, in one paper. Tenant knowledge, tools and persona are pre-encoded as cache segments and composed per call — 7.5× faster first token, 11.5× less per-tenant storage. The second half measures the wall behind it: generation on commodity compute is bound by memory bandwidth, a second model set adds no throughput, and the tool turn and the speech stream have to be scheduled around that. It is the constraint that decides how small a device can still hold a conversation.

Preprint
Also

Training write-ups for the Hindi and Spanish speech models, a speech-recognition benchmark on real call audio, and a working paper on revenue responsiveness across India’s VAT-to-SGST transition are available on request.

Education & credentials

Education and what came after.

Undergraduate

Guru Gobind Singh Indraprastha University

Bachelor of Technology, Information Technology

University School of Information & Communication Technology

Class of 2010 · 2006 – 2010
Peer-reviewed

IJCISIM, 2026

Journal publication · clinical machine learning

Intervention-aware ICU mortality prediction, with collaborators at Bennett University.

doi:10.70917/ijcisim-2026-5533
Released models

Speech, three languages

English · Hindi · Spanish

Recognition and text-to-speech trained in-house and running in production on commodity hardware.

2025 – 2026
The Inference Mechanic — book coverBook · 2026

The Inference Mechanic

Tuning LLMs for Speed, Cost, and Scale

A hands-on manual of LLM inference engineering — hardware limits, tune-ups, engines and fleets, guardrails and the edge — with every number measured on real GPUs. EPUB and PDF; the first 100 readers on the waiting list get it free.

theinferencemechanic.com · read Chapter 1 free
Enterprise record

Four institutions

GE Healthcare · S&P Global · Bristol Myers Squibb · State Street

Regulated healthcare software, market-index systems and banking platforms, from inception to production support.

2010 – 2023
Beyond work

Where the other hours go.

Court

Squash & badminton

Two or three evenings a week. Endurance from one, reflexes from the other.

Table

Ping pong

A long-standing habit that began as an office game and never quite stopped.

Outdoors

Trekking

Himalayan trails when the calendar allows. Altitude is an honest performance test.

Travel

Sightseeing

Old cities, new ones, and whatever is worth the detour between two meetings.

Reading

Geopolitics

Trade, energy, semiconductors and the states that contest them — where compute, capital and customers end up.

Reading

Technology & research

Inference-systems papers and the small-model work that lets a conversation run on a phone-sized machine.

Contact

Get in touch.

Founders, researchers, and anyone building on commodity or edge hardware — or in healthcare, finance or hospitality — are welcome to get in touch.

Emailroshan.purush@gmail.com LinkedInlinkedin.com/in/purushottam-c-377330225 Investor deckdeck.agentbox.quickdial.ai