Vamsi TalasilaTechnology & AI Advisory
Start a conversation
Work

Five case studies, told straight

Each one covers the context, the hard part, what I did personally, the result in numbers, and what I'd do differently. Clients named with permission.

Gaming & consumer internet · 2012–2021

Taking a company public at 100M-user scale

100M+users
65+countries
~6Bevents a day
99.99%uptime

Context

Nazara Technologies grew from a mobile-games company into India's first listed gaming company. As Group CTO I ran technology across 11 group companies and led a 200+ engineer organisation, with platforms delivered through 140+ telecom operators.

The hard part

Keeping platforms alive at roughly six billion events a day while the whole company was being examined for an IPO. Scale stops being a number on a slide and becomes a discipline when every outage is visible to investors.

What I did

I set the group architecture and standards, led the engineering organisation across geographies, drove M&A technology integration, and ran the technical due diligence and the investor-facing technology story. The platforms were event-driven, stateless where possible, and designed to degrade gracefully.

Result

India's first gaming-company IPO on the NSE and BSE in March 2021, on platforms serving 100M+ users at 99.99% uptime.

What I'd do differently

Standardise observability across acquired companies earlier. Integration is faster when every team sees the same dashboards from day one.

Governed agentic AI · 2023–present

An operating system for governed AI agents

257agents
23model providers
530+tools
65–75%lower run-cost

wizardstechglobal.com

Context

Companies kept building impressive GenAI demos, then spending a year unable to get them into production. The blocker was never model quality. They couldn't govern the AI, afford it at volume, or explain it to an auditor. So I founded WizardsTech Global and built WAI-SDK.

The hard part

Making governance cheap enough to leave switched on. The naive version added seconds of latency to every answer.

What I did

I architected and built the core: the gateway, the Queen Orchestrator (five decomposition strategies), the capability-based registry, routing and cascade logic, five-tier memory, and the governance and audit layer. I ran most policy checks in parallel with generation, and kept multi-model consensus for the few decisions that warrant it.

Every model call passes one governed gateway Apps & userschat, APIs, workflows Queen Orchestratorsplits work by its shape Agent registry257 agents, by capability Governed gateway policy check · PII redaction · model routing · cache · evalshuman review for high-stakes calls Small & domain models Frontier & reasoning Tools & data (530+) Audit trailhash-chained,every call loggedEU AI ActISO 42001 · DPDP Every model call passesone governed gatewayApps & userschat, APIs, workflowsQueen Orchestratorsplits work by its shapeAgent registry257 agents, by capabilityGoverned gatewaypolicy check · PII redactionmodel routing · cache · evalshuman review for high-stakes callsSmall & domain modelsFrontier & reasoningTools & data (530+)every call loggedAudit trailhash-chained, every call loggedEU AI Act · ISO 42001 · DPDP
Nothing calls a model provider directly. That single rule is what makes security, cost control and audit possible in one place, and lets a model be swapped by configuration.

Result

A production platform at 99.99% uptime, 65–75% cheaper to run, productised into a nine-product suite and several industry verticals, and earning enterprise revenue. Client engagements include AVA Connect (enterprise AI agents on AWS), IIMX (education CMS and LMS) and MediConnect.

What I'd do differently

Build the evaluation harness before the agents. I built 257 agents before a proper eval gate, and agent count quietly became a vanity metric. Now evals come first and the registry grows only as fast as each agent proves it beats a single well-prompted call.

Across platforms

AI cost engineering that doesn't cost quality

65–75%lower run-cost
50–60%LLM cost cut at QX Lab AI
99.99%uptime kept
4stacked measures

Context

Agentic systems are brutally expensive by default, because every step can burn a frontier model. At enterprise volume that makes good products unaffordable.

The hard part

Knowing when a cheap answer is good enough. Without a reliable escalation signal, a cascade either wastes money or ships bad answers.

What I did

I stacked four measures: a model cascade (small or domain model first, escalate only on a failed confidence or validation check); semantic and prefix caching; routing across 23 providers to the cheapest model that clears the quality bar; and tiered memory so agents don't re-read everything each turn.

Most requests never need the most expensive model Requestafter cache check Small / domaintried first Mid-size modelon a failed check Frontierhardest cases only escalateescalate Answer passes evals → returned, cached, audited Most requests never needthe most expensive modelRequestafter cache checkSmall / domain modeltried firstescalate on a failed checkMid-size modelsecond tryescalate againFrontier modelhardest cases onlyAnswer passes evalswhichever model passes first · cached · audited
The cascade only works with a reliable escalation signal: a confidence or validation check that says "this answer isn't good enough". Getting that signal right took several iterations of being wrong in production.

Result

65–75% lower run-cost on WAI-SDK and 50–60% lower LLM cost at QX Lab AI, where I was Chief Product & Technology Officer and helped launch AskQX in 100+ languages.

What I'd do differently

Put cost per successful outcome on the main dashboard from the start. Teams optimise what they can see.

Health & wellness · 2021–present

Preventive health at national visibility

2022Ministry of Ayush recognition
PMacknowledgement
HIPAA · DPDPaligned design
0→1platform build

aayu.live

Context

Lifestyle diseases are rising, and yoga therapy has decades of research behind it but little digital reach. I co-founded Reset Tech and, as CTO, built AAYU: AI-enabled preventive health, wellness and telehealth, with research-based programmes from S-VYASA.

The hard part

Handling sensitive health data properly on a start-up budget, and making therapy programmes feel personal at scale.

What I did

I built the platform from zero: architecture, privacy-preserving data design aligned with HIPAA, GDPR and DPDP, the AI personalisation layer, and the engineering team.

Result

In 2022 AAYU was recognised by India's Ministry of Ayush under the Start-up India 'Yoga in the Digital Era' category, and the Prime Minister acknowledged the app at the Digital Yoga Exhibition in Mysuru on International Day of Yoga.

What I'd do differently

Instrument clinical outcomes from the first user, not just engagement. Outcomes are what payers and employers buy.

Facing something similar?

Tell me the problem in a few lines. I'll tell you honestly whether I can help.

Schedule time with me →Send a brief