Shubham Pagare
shubhamp2017@gmail.com
Minimal office desk

Shubham Pagare

I turn complex AI systems into outcomes that hold in production.

About

Engineer first, fluent in the business

I'm an engineer. I build AI platforms, make systems faster and more reliable, and get that work into production. I also understand the business problem, so what ships is aimed at an outcome, not just a technical win.

Business context, not just the code Ideas → shipped reality Bridge stakeholders and engineering Make complex systems feel simple
120%
response-time gain on WebSocket bottlenecks
10×
execution speedup, Python → C++
~6 min
analysis cut from a 1-hour task
100%
tool-licensing savings vs LoadRunner
How I work

The way I think

The instincts behind the tools — how I approach systems, diagnostics, and hard problems.

Architect Mindset

See the whole system before the parts — design for scale, trace the data, anticipate where it breaks.

Deep Diagnostics

Comfortable in the weeds: thread dumps, memory leaks, latency hotspots, token-level profiling.

Problem Solving

Break ambiguous, multi-layer issues into root causes — then ship the fix that actually holds.

Stakeholder Clarity

Translate complex engineering into decisions leaders can act on.

Experience

Where I've shipped

A run through the roles, the clients, and the measurable outcomes — from AI agent platforms to deep performance forensics.

Dec 2024 — Present

Senior Performance & AI Engineer

Waynautic Technologies
Client: PwC US— AI-based product on Azure Cloud
  • Designed and deployed multi-agent AI workflows with LangChain, LangGraph, and Azure OpenAI to automate QA and business processes.
  • Built a custom E2E trace tracking tool with Langfuse for deep LLM pipeline visibility — surfacing latency hotspots, token inefficiencies, and failure points.
  • Led AI performance engineering across LLM analysis, load-testing strategy, and Azure optimization, driving measurable cost reduction.
  • Identified critical architectural bottlenecks and shaped design decisions in direct review with PwC US stakeholders.
Sept 2023 — Nov 2024

Performance Engineer

Zensoft Services
Client: PwC US
  • Achieved 120% response-time improvement and ~0.1% error rate resolving multi-layer WebSocket performance issues.
  • Built a custom JMeter + Selenium framework replacing LoadRunner TrueClient — 100% savings on tool licensing.
  • Automated test execution and analysis on Azure ADO with Python, cutting insight-delivery time by 50%.
Aug 2021 — Sept 2023

Performance Engineer

NTT Data Services
Client:UNIQLO Uniqlo (UQ)
  • Led an in-house open-source monitoring & auto-analysis tool, reducing a 1-hour task to ~6–7 minutes.
  • Diagnosed memory leaks, packet drops, and microservice issues across AWS and GCP, significantly improving production stability.
  • Built precision JMeter load models and implemented database replication for improved reliability.
  • Star Award (Foresight & Hard Work) and KK Award (Outstanding Performance of the Year).
Jan 2021 — June 2021

Engineering Intern

Softnautics LLP
  • Migrated a codebase from Python to C++, achieving a 10× improvement in execution speed.
Toolkit

What I work with

From agentic orchestration to thread-dump forensics — the stack behind the outcomes.

Generative AI & Agentic Systems
LangChain LangGraph Langfuse AI Agent Dev LLM Observability E2E Trace Analysis Prompt Engineering
Performance & Monitoring
JMeter Gatling Azure Load Testing LoadRunner Grafana Datadog Dynatrace CloudWatch App Insights
Cloud & DevOps
Azure AKS AWS ECS GCP Jenkins Azure ADO Docker Git / GitHub Microservices
Programming & Analysis
Python JavaScript React Django C++ Thread / Heap Dump AWR LLM Token Profiling
Writing

Notes from the edge

Thinking out loud on AI systems, context, and making agents useful in the real world.

Aug 20, 2026·Session·LLMs · Internals

How a language model actually reads and writes

A walkthrough of what happens when you hit send — tokens, embeddings, attention, sampling — using one sentence: “The cat sat on the mat.”

Open session
Aug 8, 2026·7 min read·AI Systems · Architecture

Why AI systems pass the demo and die in production

Most AI products don't fail because the model is weak. They fail because the architecture assumed a clean world — complete docs, linked alerts, tidy ownership — that never existed outside the slide deck.

Read essay
Aug 6, 2026·6 min read·AI Agents · MCP

Context Builder: why coding agents fail on large repos — and what to build instead

Agents don't need another chat window. They need a budgeted, task-shaped context pack — Map, Facts, Evidence — assembled ahead of the session so they stop rediscovering the monorepo every time.

Read essay
Foundations

Education & recognition

Degree

B.Tech, Computer Science (AI/ML)

MIT ADT University, Pune
CGPA 8.32/10 · Distinction · Aug 2017 — July 2021
Star Award — Foresight & Hard Work · NTT Data KK Award — Outstanding Performance of the Year · NTT Data
Certifications
2024Site Reliability Engineering — Google / Coursera
2019Data Science Orientation — IBM / Coursera
Contact

Let's build something observable.

Open to roles and collaborations in AI platform engineering, LLM observability, and performance. The fastest way to reach me is below.