Archana Turaga
I build the AI agents that 170 PayPal teams ask for help in Slack, and I run the 10+ Airflow environments they depend on.
Software engineer, Cloud Data Platform Engineering. Austin, Texas.
Full stack,
full glam.
Pick a shade.
The agent platform I built
I built it on Google's Agent Development Kit. It has been in production for a year across 50+ Slack channels, handles 25+ requests a day, and has cut support load from 18 to 9 story points a sprint.
- Someone asks in one of 50+ Slack channels.
- A classifier routes it to a model to keep token spend down. Gemini handles retrieval and Claude handles orchestration.
- The channel's domain agent takes it. There are 8+, for Airflow, Automic, Tableau, Looker and more.
- It reads live state through MCP servers, like the 40-tool Airflow one I wrote.
- It searches its RAG knowledge base.
- It answers in Slack. Resolved answers go back to Confluence to be re-indexed.
Writes such as pause, unpause and restart are guarded: the DAG's owner approves them first.
Data Platform
Channels
and 46 more
Apps
platform-agent
# airflow-supportAsk @platform-agent about DAGs, runs and failures
@platform-agent why did daily_settlement_report fail last night?
- Routed to the Airflow agent
get_dag_rundaily_settlement_reportget_task_logsload_settlementssearch_knowledge_base"column renamed upstream", via Gemini
load_settlements failed at 2:14 AM. Upstream, settlements_raw renamed fee_amt to fee_amount, so the insert couldn't find the column. The fix is a one-line mapping change, and the runbook has it.
Open runbook
Saved to Confluence for re-indexing
✅ 2
That was it. Thank you! 🙏
Message # airflow-support
platform-agent is thinking…
# looker-helpDashboards, explores and refresh questions
The Merchant Volume dashboard still shows Tuesday's numbers. Is Looker down?
- Routed to the Looker agent
get_dashboardMerchant Volumeget_dag_runmerchant_rollup
Looker is fine. The dashboard's derived table waits on the merchant_rollup DAG, and Wednesday's run is queued behind a backfill. It should land around 11:30 AM, and I'll post here when it does.
👀 1
Message # looker-help
platform-agent is thinking…
# tableau-helpWorkbooks, extracts and permissions
Which extract feeds the Risk Overview workbook, and when does it refresh?
- Routed to the Tableau agent
search_knowledge_base"Risk Overview", via Geminiget_extract_schedulerisk_daily
It's risk_daily. It refreshes every day at 6:00 AM CT, after the risk DAGs finish, and #risk-analytics owns it.
Source: Confluence, Tableau extracts
🙌 1
Message # tableau-help
platform-agent is thinking…
# dag-ownersChanges to production DAGs
@platform-agent pause merchant_rollup until tonight's migration is done
- Routed to the Airflow agent
pause_dagguarded, needs the DAG owner
Pausing is a guarded action. @Sam O., you own this DAG. Approve?
ApproveDeny
Approved
pause_dagmerchant_rollup, approved by Sam O.
Paused. Tell me when the migration is done and I'll unpause it.
Message # dag-owners
platform-agent is thinking…
I scaled it on GKE with Helm and autoscaling; it keeps state in Redis and passes work over Pub/Sub. OpenTelemetry and Langfuse handle tracing, evaluation and cost tracking. Test coverage is 98%.
An MCP server that lets agents work with Airflow safely
I designed and wrote it with FastMCP, and I wrote its clients too. Access is JWT-scoped: agents can read DAGs, runs, tasks and logs, and the only writes (pause, unpause and restart) are guarded. The agent platform uses it, and other fleet tooling has adopted it too.
The Airflow fleet I run
Open-source Airflow 2 and 3, Astronomer and Cloud Composer, running on-prem, on VMs and on GKE. Together they handle millions of task runs.
99.99%
availability for the 170 application teams that run on the fleet
10+
Airflow environments I administer, with 1,000+ DAGs in each
170
customer repos I lead through Airflow upgrades
Finding what's deprecated, rolling out in phases and fixing the DAGs that break.
300+
broken workflows flagged
by the fleet-wide DAG health scanner I built. It cleans up orphaned tasks on its own and alerts owners in Datadog.
4 yrs
primary on call, every other week
I was triage lead on a high-severity incident where a schedule misconfiguration skipped runs, stalled catch-up and delayed top-tier merchant reporting.
Provisioning and CI/CD
I automated Cloud Composer provisioning with Terraform and moved fleet CI/CD from Jenkins to Harness. I also stress-test scheduler capacity to right-size instances.
What I'm building now
- An agent pipeline that reads DAG and failure patterns and flags workflows likely to break before they reach production.
- Failure alerts that arrive with the likely root cause and a one-click rerun.
- LLM triage that runs headless, right on the scheduler nodes.
Work and school
10 years of school and work. The last 4 have been data platform and AI engineering, first at Braintree and then at PayPal.
Colored bars are jobs, gray bars are degrees, the hatched bar is the degree in progress, and dots are certifications. The dashed line is today.
Earlier projects 2019 to 2022
- 2022Diaspora
Mobile app that brings local businesses, community news and social updates together for Venezuelan immigrants.React Native, Yelp, Twitter and News APIs
- 2021Cloud deployments
Python web apps on AWS and Azure with D3.js charts, and an image recognition tool on Docker and Kubernetes.Python, MongoDB, SQL, D3.js
- 2021University management system
Web app with login and student records.PHP, Laravel, SQL
- 2020Heart disease prediction
KNN classifier over 13 clinical features, checked with k-fold cross-validation.Python, KNN
- 2020Sound classification
CNN that classifies audio events, aimed at security-relevant sounds.Python, CNN
- 2020COVID-19 outcomes
Linear regression relating ICU admissions to deaths.Python, regression
- 2019Fruit nutrition
CNN that recognizes fruit in a photo and returns its nutrient profile.Python, CNN
What I build with
- Agents
- Google ADK
- MCP and FastMCP
- LangGraph and LangChain
- LiteLLM
- Pydantic
- Gemini and Claude
- Langfuse
- Platform
- Apache Airflow 2 and 3
- Astronomer
- Cloud Composer
- Automic
- GKE and Kubernetes
- Helm
- Terraform
- Data
- BigQuery
- Bigtable
- Pub/Sub
- Kafka
- Redis
- PostgreSQL
- Redshift
- Code and ops
- Python
- SQL
- FastAPI
- pytest
- OpenTelemetry
- Datadog
- Harness
Moved off
-
Jenkins to Harness
I moved the Airflow fleet's CI/CD.
-
Redshift to S3
Braintree's exit migration. I built the operators for it.
Makeup is my other stack
Off the clock, I paint in oils. I trained at the Siri Institute of Painting, finishing the senior course in 2016 and the advanced course in 2018.
The lipsticks at the top of this page set its colors, and the line you followed down here is drawn in the one you picked.
Write to me about agents, MCP or Airflow
LinkedIn (opens in a new tab)Austin, Texas