Skip to content
View ana-lan's full-sized avatar
🙃
:)
🙃
:)
15 followers · 17 following

Achievements

Achievement: Pull Sharkx2

Achievements

Achievement: Pull Sharkx2

Block or report ana-lan

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ana-lan/README.md

Anagha Langhe

ML / Data Engineer — building RAG systems, data pipelines, and poking at what's inside LLMs

MS CS @ UT Dallas (May 2026)  ·  Dallas, TX

LinkedIn Gmail GitHub LeetCode


Currently

  • 🔍 Digging into mechanistic interpretability — trained an SAE on GPT-2-small and built a live safety monitor from it
  • 🛠️ Shipped a RAG evaluation tool at an OpenAI hackathon (Build Week)
  • 💼 Past experience across ML (IoT/CV at MIT-WPU × Capgemini), data engineering (DRDO — LiDAR/camera perception pipelines), and TA'ing DSA at UT Dallas
  • 🎓 MS in Computer Science from UT Dallas, open to full-time ML / Data Engineering roles

Featured Projects

Downshift  ·  live demo  ·  PyPI Cuts LLM costs per PR. Finds every LLM call in a Python repo, tests cheaper models against evals written for each call site, recommends the safe downgrades, and posts the projected monthly cost change on every pull request. On the demo app, 3 of 8 call sites downgraded for an 18% projected saving with pass rate held steady.

Python AST Analysis LLM Evals GitHub Actions IBM Bob

ActivationLens Mechanistic interpretability on GPT-2-small — trained a sparse autoencoder on the layer-6 residual stream, cut dead-feature collapse from ~80% to 0.77%, then built a live per-token safety monitor (0.758 AUROC) with a one-pass kernel that trimmed monitoring overhead 6x.

PyTorch TransformerLens SAE FastAPI Next.js

OpenAI Build Week GPT-5.6-powered tool that scores, diagnoses, and auto-tunes RAG pipeline outputs — built with Codex during OpenAI's Build Week hackathon.

Python OpenAI API RAG

DocumentSync AI RAG pipeline over 30 ArXiv papers across 4 chunking strategies, evaluated with RAGAS across 120 LLM-as-judge calls — best config improved context precision by 59% and recall 3x over baseline.

LangChain ChromaDB Sentence-Transformers Streamlit

Real-Time Crypto Streaming Pipeline Streams live market data via Kafka + Spark Structured Streaming into a Redshift star schema, orchestrated with Airflow — cut Athena bytes scanned by 89%.

Kafka Spark Airflow AWS

Stack

Languages
Python Java C++ JavaScript SQL

ML / AI
PyTorch TensorFlow Keras Scikit-Learn ONNX Runtime NumPy Pandas OpenCV Hugging Face LangChain RAG GenAI LLM Mechanistic Interpretability Sparse Autoencoders TransformerLens Sentence Transformers Prompt Engineering

Data Engineering
Apache Spark Apache Kafka Apache Airflow AWS Glue AWS Lambda Hadoop Databricks dbt Hive Data Warehousing

Backend & Web
Node.js Express React HTML5 CSS3 Docker Redis Postman JWT REST APIs Microservices System Design Distributed Systems CI/CD OOP

Cloud & Databases
AWS Amazon S3 Redshift Athena QuickSight CloudWatch MySQL MongoDB

Analytics & Visualization
EDA Data Visualization Tableau Power BI Streamlit

Tools & Robotics
Git GitHub Linux ROS


Snake animation

Pinned Loading

  1. activation-lens activation-lens Public

    Mechanistic interpretability × inference research on GPT-2-small — a sparse autoencoder benchmarked against quantization, converted into a live low-overhead safety monitor.

    Python

  2. downshift downshift Public

    Cut LLM costs per PR. Finds every LLM call in a Python repo, proves which can safely run on cheaper models with per-call-site evals, and flags cost increases in every pull request. Built with IBM Bob.

    Python

  3. openai-build-week openai-build-week Public

    GPT-5.6-powered tool that scores, diagnoses, and auto-tunes RAG pipeline outputs — built with Codex for OpenAI Build Week

    Python

  4. real-time-crypto-data-pipeline real-time-crypto-data-pipeline Public

    End-to-end real-time crypto data pipeline

    Python

  5. DocumentSyncAI-RAG-based-QA-for-Documents DocumentSyncAI-RAG-based-QA-for-Documents Public

    RAG pipeline for document Q&A — 4 chunking strategies evaluated with RAGAS metrics, Streamlit UI

    Python