Pulin Pathneja
Sr. Director, AI Engineering · Freshworks
I design, ship and scale autonomous multi-agent systems in production: LLM agents, self-optimising Martech and predictive operations, serving 350M+ users on a platform processing 3.2 trillion records a day.
Agent topology
7 production agents around one orchestration core
Platform
Data and throughput behind the agentsDaily records processed
Trillion records per day, 2017–2025
Team by vertical
45+ engineers, share of headcount
Peak throughput by system
Transactions per second
Agents
7 deployed in production| Agent | Organisation | What it does | Headline | Status |
|---|
Martech agent loop
How the flagship platform works
Event stream
Sample of the kind of events the agents emit
Anatomy of a real-time decision
Illustrative trace of one Martech decision inside the <60ms budget
Impact
Before and after the agentsNetwork RCA agent
Complaint resolution, before → after
Outcome gauges
Headline results across agents
Career
Roles timeline
BE (IT), Punjab Engineering College, 2001–2005
Focus over time
Relative share of work by theme, per year (self-assessed)
Skills
Self-assessed proficiencyCompetency map
Score out of 100
Stack
What I build with
Agentic AI
- Multi-agent orchestration
- LLM agents
- Tool-use
- RAG pipelines
- Human-in-the-loop
- Guardrails
- Autonomous workflows
AI & data leadership
- AI strategy
- MLOps / LLMOps
- GenAI
- NLP
- Computer vision
- Data monetisation
- Product strategy
Infrastructure
- Python
- GCP
- AWS
- BigQuery
- Kafka
- Spark
- Airflow
- Docker
- Kubernetes
- LangChain
- Vector DBs
What people say
Writing
All posts on LinkedIn ↗Everyone's talking about agentic AI. Almost no one is deploying it.
Multi-agent inquiries are up 1,445%, yet only 11% of companies run agents in production. The model is 5% of the system; orchestration, guardrails, tools and observability are where companies fail.
Read on LinkedIn → ArchitectureDon't build a smarter agent. Build a smarter system of agents.
Five years of deploying agents at 350M-user scale: a Planner, Executor and Evaluator loop beat any single "brilliant" agent. 100K+ TPS, sub-60ms, +15% engagement.
Read on LinkedIn → Case studyFrom 30 days to 3: resolving network complaints with RCA agents.
Agents that diagnose root causes across millions of signals, prioritise by severity and trigger fixes. Accuracy went from 65% to 89%, without adding headcount.
Read on LinkedIn →Let's talk agents.
Alongside building digital workers at Freshworks, I advise founders and leadership teams on agentic AI strategy, multi-agent architecture and data platforms.