tharun
The day job is data science at AT&T. The night project is an agent that fills out job applications on its own — and knows exactly when to wake me up.
What brings you here?
or just scroll — I'll narrate as you go
off duty · zero bars of signal
tharun> status --now
A portfolio is a snapshot; this is the live feed.
The job application agent below — currently wiring up the Gmail confirmation watcher and hardening the checkpoint flow.
Deep in an 8-week agentic AI sprint: LangGraph, CrewAI, and LlamaIndex courses, plus Google Cloud Skills Boost labs on Vertex AI and Terraform.
Anthropic's engineering posts on building effective agents. Strong opinions on when not to use one.
tharun> projects --order-by shipped_last
Nine public repos below, plus the work-in-progress I can't stop tinkering with.
A semi-autonomous agent that takes a job posting from URL to submitted application. LangGraph runs the workflow, Playwright scrapes and fills forms, Postgres remembers every application, AWS Secrets Manager keeps credentials out of the code, and the Gmail API watches for confirmations.
The interesting part isn't the automation — it's the restraint. The agent stops at exactly two points and refuses to continue without me:
LangGraph · Playwright · Postgres · AWS Secrets Manager · Gmail API · Streamlit
Ask a question in English, get an answer backed by SQL. An Anthropic SDK tool-use loop that writes queries, runs them, reads results, and keeps going until the question is actually answered.
Anthropic SDK · SQLite · Streamlit view source →Document question-answering with retrieval done right — LangChain orchestration, ChromaDB vector store, answers grounded in retrieved context instead of vibes.
LangChain · ChromaDB · embeddings view source →Predicting one-year post-transplant outcomes — death, graft failure, re-transplantation — with Cox PH, logistic regression, and random forests. Medicine keeps the statistics honest.
Cox PH · random forest · survival analysis view source →30-day readmission risk with SHAP explainability, so a clinician can see why a patient scores high — served as a Streamlit app.
scikit-learn · SHAP · Streamlit view source →Cox proportional hazards on clinical data with an interactive dashboard for hazard ratios, survival curves, and covariate effects.
Cox PH · Streamlit · statistics view source →Classification on heavily imbalanced transaction data — the classic problem where accuracy lies to you and precision-recall tradeoffs decide if the model is useful.
XGBoost · imbalanced learning view source →Hybrid LSTM + statistical forecasting for operational metrics, with anomaly detection and seasonality decomposition.
LSTM · ARIMA · Prophet view source →Feature engineering, model comparison, and experiment tracking wired into one reproducible pipeline. One command, not a folder of notebooks.
Python · experiment tracking view source →Advanced SQL on NYC Taxi data, written the way analysis actually happens: business question, query, finding, recommendation.
SQL · window functions · Python viz view source →tharun> experience --reverse-chronological
Telecom scale, academic medicine, enterprise consulting. Three very different data cultures; the constant is shipping work people act on.
AT&T
SEP 2024 — PRESENT
San Francisco, CA
Stanford Medicine
SEP 2023 — AUG 2024
Redwood City, CA
LTIMindtree
JAN 2022 — JUL 2022
Hyderabad, India
tharun> skills --group-by domain
Colored pills are what I reach for most right now.
tharun> education
Statistical Analysis · Machine Learning · Deep Learning · Time-Series Analysis · Causal Inference · Bayesian Methods
Also tutored statistics through SFSU's SSS-TRiO program — the best test of whether you understand regression is explaining it to someone seeing it for the first time.
tharun> contact --channel any
Open to Data Scientist, ML Engineer, and applied AI roles. If you're building something where production ML meets agents, I want to hear about it.