Jay Gopalan, engraved. Scroll and it turns to 7,500 dots.
JayGopalan
Rent the genius, own the competence.
If you can’t prove it, don’t ship it.
15+ years shipping machine learning, deep learning and generative AI to millions of people, and 29 years building systems for banks and brokerages. Today I work on voice AI architecture and solutions at Netomi, from Dallas.
Keep scrolling. The same dots tell the rest.
Gopalan’s Laws
First law
At scale, rare is routine.
At a million answers a day, 99.9% right is still a thousand wrong ones, every day.
Second law
The cost of an AI answer falls 10x a year. The cost of a wrong one never does.
Netomi, 2025 to now
Forty million phone calls a year, answered by voice AI agents.
Voice AI architecture and solutions at Netomi: agents running at more than 6,000 concurrent calls and handling over 40 million calls a year for enterprise clients. Routine calls get resolved end to end, so clients run high-volume support with a small human team and save millions a year.
The work spans telephony, cascaded and speech-to-speech pipelines, LLM orchestration, integrations with systems of record, small models fine-tuned for voice, and the guardrails and evaluation that keep agents dependable in regulated environments.
In voice AI, every second of silence counts. Exponentially.
msg.ai, now Netomi, 2015 to 2019
From five engineers to seventy.
As Vice President of Technology I grew engineering across the US and India and built the delivery organization behind enterprise conversational AI. I sat on the leadership team that set the product vision while the company's valuation passed $150M.
The deep-learning platform we built (language understanding, intent classification, dialogue management, real-time pipelines) automated customer service for enterprise clients and served millions of end users.
Wall Street, 2003 to 2023
A career spent where the money moves.
Goldman Sachs, JPMorgan Chase, UBS, Merrill Lynch, Lehman Brothers and Barclays, Citi, BNY Mellon. I built low-latency, high-throughput systems for prime brokerage, settlements and asset servicing, and moved mainframe workloads to modern platforms. One revamp cut server use by 75%; one migration saved over $2M a year.
Then I crossed to the other side of the desk. I co-founded Fyniti and S&L Capital Markets, took the broker-dealer through SEC registration and FINRA membership, and served as its General Securities Principal. We launched three products and sold the AI equity-research and basket-trading businesses.
Research
Quantum computing, written down and filed.
The quantum research program I started at Fyniti, applying quantum algorithms and machine learning to trading, became a three-volume book series and five US provisional patent applications, filed in April 2026 with me as sole inventor. They cover correlation discovery, arbitrage detection, quantum search and optimization.
The remaining business became a quantum computing company. I still advise its team.
Quantum hype stays in superposition until you benchmark it.
NoblQ, 2023 to 2025
Small models, built to be owned.
I built NoblQ's AI Center of Excellence from nothing: strategy, governance, and an "AI-first" program for the whole firm. The accelerators we shipped raised engineering productivity by about 70%.
I fine-tune small, domain-specific models that give clients a private, lower-cost alternative to frontier APIs, and I publish open models on Hugging Face. The agent harnesses I wrote (orchestration, tool use, evaluation, observability) now sit under voice agents, compliance document processing, and AI training platforms for universities.
That's the short version.
The books and the full record are below. If you're working on enterprise AI, voice agents, sovereign or private models, or quantum methods in finance, I'd like to hear about it.
A pair from the Taming Your AI series: one book for the people who decide, one for the people who build. Coming soon. Read chapter 1 at tamingyourai.com
Patents. Five US provisional applications, filed April 2026, sole inventor: quantum and AI methods for correlation discovery, arbitrage detection, quantum search and optimization in trading.
White papers. Quantum tensor networks for basket trading. Quantum arbitrage detection. Quantum-accelerated credit risk.
Licenses. FINRA Series 24 and Series 7, NASAA Series 63. View on BrokerCheck.
Voice AI agents at 6,000+ concurrent calls and 40M+ calls a year for enterprise clients.
NoblQ
Head of Center of Excellence, AI
Built the AI CoE, shipped accelerators that lifted engineering productivity about 70%, and owned security and data integrity.
Fyniti Global Equities and S&L Capital Markets
Co-founder; General Securities Principal
SEC-registered, FINRA-member broker-dealer. Tokenization platform, Fyniti IQ and Fyniti Blocks.
msg.ai (now Netomi)
Vice President of Technology
Grew engineering from 5 to 70 across the US and India; built the deep-learning platform.
BNY Mellon
Principal Architect, Vice President
Fastest data-processing stack in Asset Servicing at the time; redesigned and migrated the Enterprise Reporting System.
Citigroup, Prime Finance
Vice President
Event orchestration on a complex-event-processing engine, 30% faster. Cut resource use 75%, saving $1M+ in servers.
Lehman Brothers and Barclays
Technology Lead
Settlement, repo and position feeds through the consolidation; mainframe migration saving $2M+ a year.
Merrill Lynch, Global Prime Brokerage
Lead Developer
Designed the core of the prime brokerage reporting platform.
UBS
Lead Developer
Financial-control workflow redesign, 40% faster.
JPMorgan Chase
Lead Developer
Moved BankOne applications onto a modern stack after the merger.
Goldman Sachs
Senior Developer
Client account opening systems.
AdventNet (Zoho), eSpace, Crompton Greaves
Early career
JMX 1.2 core, and the trading engine behind HomeTrade, one of India's first real-time online trading portals.
ழ
ழகரம்
Zhagaram Labs
An independent AI research lab (non-profit) for the Tamil language.
I founded it and contribute to it, with a group of volunteers. Tamil is my mother tongue, and AI still serves it poorly. We build speech and language models for the more than 80 million people who speak Tamil, starting with voice, because many of them will talk to AI long before they type to it. Most of what we build is open source.
67%of Tamil words OpenAI’s Whisper gets wrong in read speech, measured on Uraikal, the lab’s open benchmark.
ழ, zha, is the last sound in தமிழ், and the one speech models most often get wrong. One letter apart:
வாழ்
live
வாள்
sword
வால்
tail
Kural குரல்A voice model you talk to in everyday Tamil, Tanglish and regional speech. In progress.
UraikalThe open benchmark: spoken Tamil, dialects, Tanglish and ழ ள ல. Nothing ships unless it beats the last version.
Sevi செவிSpeech recognition for every dialect and noisy phone audio. In progress, open weights.
Oosai ஓசைSpeech synthesis that sounds like spoken Tamil, not a newsreader. In progress, open weights.