Retrieval that cites its sources
Our own crawlers, Airflow fanning chunking out across workers, Nomic embeddings and Elasticsearch as the vector store, so answers are grounded in real product knowledge.
answer.cite(context)
▲ 17.38°N 78.48°E · Hyderabad, India
Hi, I'm
Muthya Varshith Vankamamidi · Senior Software Engineer, Agentic AI @ Workato
I build AI agents for customer support at Workato. Six years in and still climbing.
git blame hello.md
a7e3f91now1# hey, I'm Varsh
2
deadbeesince forever3Friendly neighbourhood tech enthusiast. If
deadbeesince forever4there's a new gadget, I've already taken it
deadbeesince forever5apart (and mostly put it back together).
6
c5f4cad10 pm · gaming7Most nights I'm deep in a Counter-Strike match
c5f4cad10 pm · gaming8that I sometimes win. // win rate: classified
9
88480ffweekends · hiking10Weekends belong to the mountains: a trail, a view,
88480ffweekends · hiking11good friends, and I'm happy.
b0a4d1eany excuse · travel12If it involves a boarding pass, count me in.
13
f00d1e5sundays · kitchen14Master chef, at least in my own kitchen:
f00d1e5sundays · kitchen15loud music, too many pans, and pretending
f00d1e5sundays · kitchen16I'm on a cooking show.
17
0ddba116 am, rarely18I also run. Lazily. Very lazily.
19
be3fca7fridays20Motto: work hard, party harder.
be3fca7fridays21And yes, ranting on Reddit counts as cardio.
22
c4d82e09 to 5 · Workato23By day I build AI agents at Workato. The rest
c4d82e09 to 5 · Workato24of this page is about that part.
26
anatomy of a support agent
What I build day to day: a ticket comes in, the right knowledge is found, an agent reasons and acts, and a human gets pulled in only when it matters.
A Python class named Varshith that extends an AI Agent: it retrieves context from a support knowledge base, reasons with an LLM, hands off to a human when confidence is low, and otherwise sends a cited answer.
⎇ main
Ln 1, Col 1
Python 3.12
UTF-8
AI online
Agents that know when to stop
Agents whose skills are Workato recipes and connectors, with a flow engine between the agent and the decision tree that hands off to a human agent when the AI can't answer.
if conf < 0.85: handoff()
Safe at a million chats a day
A PII service for seven Asian languages: LoRA-tuned classifiers plus rule-based checks, 99.9998% accurate across 2M conversations, with guardrails on every model upgrade.
redact(ticket, lang="ja")
git log --career
Six years, one long climb. Every role is a commit, newest first; open one for the full diff.
▲ elevation profile
2018 → now · click a camp to jump
-
Senior Software Engineer, Agentic AI
Building agentic AI for customer support: a PII redaction service for seven Asian languages, an event-driven voice & chat pipeline, and the Zendesk Copilot.
$500K+yearly revenue from the Zendesk Copilot7 languagescustomer data auto-redacted, near-perfect accuracy10,000+requests per second, per serviceshow detailshide details· git show a7e3f91
- Built a PII redaction service for Japanese, Korean, Chinese and four more Asian languages, on a platform handling about a million conversations a day per customer. LoRA-fine-tuned classifiers plus regex and rule-based checks reached 99.9998% accuracy across 2M test conversations. A multi-million-dollar project run by five engineers.
- Built an event-driven voice and chat pipeline: pre-processing, redaction, post-processing, summarization and per-customer tasks across nine services, each handling 10,000+ requests per second.
- Built the Zendesk Copilot, which summarizes the ticket, pulls the knowledge agents need, surfaces similar past resolutions and flags live incidents. It earns $500K+ in ARR on its own.
- Built a Salesforce handover service (MIAW), integrated MCP servers and Workato Genies, and moved our LLMs to newer models behind guardrails, cutting costs.
-
Software Engineer
Lead engineer of AgentX Support, Workato's AI support platform: AI agents, a RAG knowledge pipeline, human handoff and analytics, built from scratch.
6 monthsfrom idea to launchLeadowned infrastructure and engineeringshow detailshide details· git show c4d82e0
- Co-architected AgentX Support with our VP: AI agents whose skills are Workato recipes and connectors.
- Designed the flow engine between the AI agent and the decision tree, which hands a conversation to a human agent when the AI can't answer.
- Built the RAG pipeline: our own crawlers, Airflow fanning out chunking, Nomic embeddings and Elasticsearch as the vector store.
- Built search and agent analytics on Kafka, ClickHouse and Holistics, and kept legacy DeepConverse customers running through support and security fixes.
-
Software Engineer · DeepConverse (acquired by Workato)
Joined right before the acquisition and led the migration of DeepConverse's entire support platform into Workato with Terraform and Kubernetes.
32microservices moved over~1Mconversations a day, kept runningshow detailshide details· git show 8f1b7a3
- Joined specifically to learn the platform inside out, so I could own it once it moved to Workato.
- Migrated 32 microservices and four databases (MongoDB, PostgreSQL, Elasticsearch, Redshift) with Terraform: first onto AWS EKS, then gradually onto Workato's own Kubernetes clusters, before handing it to the infra team.
-
Software Engineer II
Built the event-driven testing platform that let the whole company ship faster, and Phoenix, which repairs broken UI tests on its own.
1 week → hoursQA time for big releases85%of broken UI tests fixed automatically38% fewerflaky testsshow detailshide details· git show 5c09b2d
- Regression platform on Terraform, RabbitMQ, Argo CD and GitHub Actions: every change gets its own short-lived environment and database, and only the affected tests run, in parallel.
- Replaced Jenkins jobs that took days and often got stuck, making test automation 60% faster.
- Phoenix re-finds UI elements after front-end changes by matching nearby, similar elements. The team later added LLMs to reach 98%; it's still in production.
-
Software Engineer
Built the Slack bots and infrastructure-as-code tools every team used to deploy and test, and won a Kudos Award.
~300engineers using my toolsdays → minutesto rebuild a test environmentshow detailshide details· git show e41a8c7
- Slack bots to deploy, test, see results and re-run pipelines, without a separate VPN for every environment.
- Terraform-based tooling across 8 products, each with 10–15 microservices.
- Elasticsearch and Kibana logging for test environments, with pipeline failures alerted in Slack.
- Built a video-recording framework that generated data for ML plagiarism detection. Kudos Award
-
Software Engineer Intern
Joined through the Winter Interns 2022 program, worked on reliability engineering on AWS and Kubernetes, and converted to full-time.
-
Full Stack Developer
Built a legal search product end to end: users find the past civil court rulings most relevant to their case, served by a cell-based, serverless backend on AWS.
End to endweb app, APIs and AWS infrastructureServerlesscell-based architecture on AWS Lambdashow detailshide details· git show 9b2c6a0
- Built search over a knowledge base of previously ruled civil court cases, hosted on Elasticsearch, so users could find the rulings closest to their own case.
- Designed the backend as a cell-based architecture on AWS, with Lambda functions, DynamoDB and other managed services, so cells scale and fail independently.
- Owned the full stack, from the web app to the APIs and infrastructure.
⎇ early-days
2018 – 2021 · where it started
Freelance Developer
Self-employed
Flask and Django apps for clients, with APIs, auth and payments, while still in college.
B.Tech, Computer Science
NIIT University
Led the tech team for siNUsoid, the university's science and tech fest.
tree ~/stack
▾ ai/
Agentic AIAI agentsRAGLLM guardrails
MCPWorkato GeniesLoRA fine-tuningPII detection
Embeddings · NomicVector searchReranking
Prompt engineeringOpenAI
▾ backend/
PythonRubyFastAPIDjangoFlask
GroovyShellREST API designMicroservices
Workato recipesSalesforce MIAWSeleniumSlack bots
▾ infra/
AWSLambdaDockerKubernetes
TerraformArgo CDRabbitMQJenkins
GitHub ActionsEvent-driven architecture
▾ data/
ElasticsearchPostgreSQLMongoDBRedshift
KafkaClickHouseHolisticsAirflowMySQLKibana
ls ~/certificates
Awards and certifications. Click one to open it full size.
Kudos Award
HackerRank employee award for building a video-recording framework that generated data for ML-based plagiarism detection.
ssh guest@varshith
Prefer the command line? Type help. Press ⌘K or / anywhere to jump here.