Vishalan Gharat

Vishalan Gharat

Senior Engineering Manager, AI

BrowserStack | Ex-Morgan Stanley | Ex-Oracle | IIM Ahmedabad (AGMP)

I lead applied AI engineering at BrowserStack, taking LLM agents from prototype to production: visual review, browser automation, root-cause analysis, and our MCP server. My work sits where models meet the business: evals, cost per task, reliability and AI governance. Before AI, I spent a decade on real-time streaming, core banking at Oracle and wealth-management data at Morgan Stanley, with a management grounding from IIM Ahmedabad.

Off-hours: FIFA, 3D printing, embedded hardware, and content automation.

Download resume

Key AI Projects

Projects I led or owned. Credit is noted where my team or the platform team built it, and outcomes are ratios rather than internal figures.

Agent-ready platform and MCP (AI Transformation)

The AI products needed one agent-ready foundation, and external AI assistants needed a clean way to drive BrowserStack.

  • Set the AI Transformation direction for the AI Features org.
  • The platform team built the agent-ready platform: a shared inference loop, guardrails, memory and an SDK.
  • Took over and grew the MCP server to 18+ tools across 4 products, with remote MCP (OAuth and one-click install).
18+ MCP tools, 4 productsRemote MCP (OAuth)~75% of MCP code AI-generated

Percy Visual Review Agent

Visual regression testing left teams triaging piles of diffs by hand.

  • I owned the lane; my team built a multimodal LLM agent that classifies and explains visual diffs.
  • Ran an eval-driven model bake-off (82% coverage / 85% accuracy for the chosen model, versus 76% / 75%).
  • Drove a 10% to 100% canary rollout and a provider re-platforming.
GA Oct 2025Inference cost ~75–90% lowerAdoption 86% → 91%Consistency 50% → 89%

Natural-language browser test agent

Web test automation written in plain language that stays stable as the web under it changes.

  • Drove the pivot to a visually grounded agent built on the open-source Magnitude framework.
  • Validated on an internal testbed scored with an LLM judge.
Step stability 85–87% → 98%First action 25s → 14s3-month retention 61% → 79%India data residency

LLM platform review and gateway

Product surfaces reached models through separate, mostly ungoverned paths.

  • Reviewed the LLM pipelines and model access, and found five access paths with only one governed.
  • Benchmarked the two main pipelines (82% versus 69% recording accuracy, about a 7x gap in P90 cost per test) to set the mandate.
  • Prototyped a single governed, model-agnostic gateway.
5 access paths, 1 governed~7x P90 cost gap

These figures are pre-gateway benchmarks that set the mandate, not gateway results.

Test Management AI

Pick the tests worth running, and turn PRDs into test cases.

  • Semantic test selection with embeddings, pgvector and a weighted rule engine, shipped as a Jira app.
  • An eval-first pipeline from PRD to test cases.
  • Took this and three other initiatives through the company's internal patent review. None were ultimately filed, but the process taught me how patentability is judged, from prior art and novelty to claim drafting.
Coverage 83.4% → 90.4%Shipped Oct 2024

Experience

BrowserStack

Mumbai, India (Remote) · Dec 2019 – Present

Senior Engineering Manager, AI · Apr 2026 – Present
  • Lead AI engineering for the AI Features org across Percy, Salesforce test automation, browser test automation and the platform squad.
  • Led the assessment and first-phase merger of the accessibility rule engine into the AI org.
  • Own LLM budget and multi-provider strategy across Anthropic, OpenAI, Google Vertex AI and AWS Bedrock, including revenue per AI dollar by product.
  • Built an AI spend console: unified spend across providers, anomaly digests, queryable over MCP.
  • Led ISO/IEC 42001 AI management system certification, passed Jun 2026.
  • Authored the agentic harness spec: memory, agent identity, guardrails, an MCP contract registry, deny-by-default writes and a model-agnostic gateway.
Engineering Manager, AI Platform & AI Features · Oct 2023 – Mar 2026
  • Built the AI squads' operating model and production-readiness gates for AI GA.
  • Helped scale the AI engineering squads across multiple teams.
  • My team shipped the Percy Visual Review Agent, a browser test agent and a root-cause-analysis agent; I took over and grew the MCP server.
  • LLMOps: provider fallback, rate limiting and event-scale capacity planning.
  • AI security: SOC 2 for the AI platform, the first internal AI pentest (OWASP LLM Top 10) and Zero Trust.
  • Set the AI org's first engineering OKRs and ran an AI-assisted engineering program (25.6% average efficiency gain).
Senior Lead / Team Lead, Live · Apr 2021 – Sep 2023
  • Led the squad owning Live and the shared MySQL fleet.
  • Org-wide alert routing cut alerts 85% and MTTA from about 35h to 1h40m.
  • Ran the MySQL 5.7 upgrade benchmark.
  • Gave a large US bank browser-based Mac access within a day, and four dedicated Macs within four weeks.
Software Engineer, Live (real-device streaming) · Dec 2019 – Apr 2021
  • WebRTC / nodeRTC streaming: cut connect time 12%.
  • TURN high availability.
  • Owned Live Ops.

Morgan Stanley

Mumbai, India · Jan 2019 – Jan 2020

Senior Associate, Data Strategy (Private Wealth Management) · Jan 2019 – Jan 2020
  • Built an Apache Ignite data fabric over 100M+ records; cut API latency 1.7s to 0.5s.
  • Built a sales-orchestration dashboard that automated 95% of manual activity.

Oracle

Mumbai, India · Aug 2016 – Dec 2018

Application Developer II (Oracle Banking Platform) · Aug 2016 – Dec 2018
  • Built a decision-tree credit-scoring model (86% accuracy) that cut third-party API cost 50%.
  • Built a pluggable REST framework serving 3,000+ banking services.
  • Built a credit-monitoring dashboard over 1B+ data points.

How I lead AI teams

  • Operating model & readiness gates. Every AI feature clears the same production-readiness bar before GA.
  • Eval-first delivery. Model bake-offs, testbeds and LLM-as-judge before anything reaches a user.
  • LLM cost discipline. Multi-provider strategy, fallback, and spend I can see and query per product.
  • AI governance. ISO/IEC 42001, SOC 2 and OWASP LLM Top 10 built into how the org works, not bolted on.

Skills

LLM apps & agents

LLMsAgentic workflowsTool callingMCPRAGEmbeddings / pgvectorMultimodal modelsPrompt engineeringBrowser agents

Evaluation & LLMOps

Eval testbedsModel bake-offsCanary rolloutsGuardrailsPrompt tracingProvider fallbackCost & latency optimization

Models & platforms

Anthropic ClaudeOpenAIGoogle Gemini / Vertex AIAWS BedrockAzure OpenAIOpen-weight SLMs

AI governance & security

ISO/IEC 42001SOC 2OWASP LLM & MCP Top 10Zero TrustAI risk assessmentData residency

Engineering

PythonJavaJavaScript / Node.jsRuby on RailsSpring BootAWSKubernetesKafkaRedisMySQLPostgreSQLApache IgniteWebRTC

Leadership

AI strategyOrg design & scalingHiringOKRsLLM budget & vendor strategyAI-assisted engineering

Writing

  • I Took Seventeen Reels Apart, Frame by Frame

    I Took Seventeen Reels Apart, Frame by Frame

    October 30, 2026

    I downloaded seventeen popular explainer reels, cut them into frames, transcribed them and measured their audio. The patterns were specific enough to code: a number that moves, a...

  • A Content Studio That Runs Without Me

    A Content Studio That Runs Without Me

    October 18, 2026

    A single RTX 3090 in my house runs an unattended content pipeline: it finds a story, writes a script with claude -p, synthesizes a voice, scores a music...

  • Cult Focus: A Coin You Tap to Lock Your Phone

    Cult Focus: A Coin You Tap to Lock Your Phone

    October 16, 2026

    Screen Time fails because the same hand that sets the limit lifts it. So the unlock became a coin you tap. The honest version: a free app and...

All writing

Education & certifications

2023 – 2024

Accelerated General Management Programme

IIM Ahmedabad

A one-year programme for working professionals covering the functions of business management plus contemporary, new-age topics, broadening the business perspective and the tools for managing modern businesses.

2012 – 2016

B.E., Computer Science & Engineering

University of Mumbai

8.09 CGPA.

2018

Machine Learning

Stanford Online, Coursera

The widely-taken Machine Learning course from Stanford Online. Certificate.

2018

Google India Challenge Scholarship (Android)

Google · Udacity

Android development scholarship sponsored by Google on the Udacity platform.

Earlier work

Shipped apps, startups & college projects

Cult Focus

A focus app paired with a 3D-printed NFC coin you tap to lock your phone, from my store The Cult Project. Native on every platform: iOS, Mac, Android.

Homely

Software intern at a growing food startup (2016). Built an Android app for order tracking, product management, analytics and billing, and the company's single-page site. Play Store.

SmartGrid (Android app)

Final-year project: a messenger that pools the compute of phones on a wireless grid to solve a shared problem faster. Won 1st prize (₹70,000) at a national IEEE project competition. Play Store.

Koinex Monitor (Android app)

Price monitoring and alerts for the Koinex crypto exchange. Play Store.

E-Yantra Robotics (IIT Bombay)

Team leader; 5th nationally across 6,000+ teams in a research-and-build robotics competition, building a cargo-alignment robot and a weed-plucking robot.

CSI-Cryptex & E-Cell, FCRIT

President of the CSI-Cryptex workshop committee and operations head of the college Entrepreneurship Cell (2014–2015); ran Android development workshops.

StupidSid

Recognised technical content writer for the student resource site stupidsid.com (2015).

More projects

School-bus tracking app, a college intranet app and site, and mindassets.in, a job portal for differently-abled individuals.

Contact

Location
Mumbai, India (Remote)
Email
vishalangharat@gmail.com
© 2026 Vishalan Gharat