Anagha Langhe // terminal

~$ ls experience/

Experience

teaching_assistant.log Data Structures / Algorithms

Teaching Assistant – Data Structures & Algorithms

The University of Texas at Dallas

Sep 2025 → May 2026

JavaPythonC++ Data StructuresAlgorithms

Mentored 100+ students through code reviews in Java, Python and C++, coaching test-case design for trees, graphs, and dynamic programming.
Redesigned grading rubrics for 3 courses, standardizing evaluation of OOP principles and time complexity analysis.

drdo_intern.log ML / Computer Vision / Robotics

Data & Machine Learning Intern

Defence Research & Development Organisation (DRDO), India

Jul 2023 → Dec 2023

C++PythonROS TensorFlowPyTorchOpenCV Sensor FusionLiDAR

Built end-to-end ML inference pipeline for multimodal LiDAR-camera perception on KITTI — sub-80ms latency per frame, 40% faster experiment turnaround across 15+ runs.
Engineered LiDAR-camera calibration module using ICP and OpenCV, reducing reprojection errors by 35%.
Enhanced object detection accuracy by 25% on KITTI; GLENet-VR outperformed baseline LiDAR approaches by 15%.
Automated data validation and regression pipelines, achieving 50% reduction in evaluation loss on the benchmark.

capgemini_intern.log ML / Computer Vision / IoT

Machine Learning Intern

MIT World Peace University × Capgemini

Jul 2022 → Jun 2023

CNNSVMTensorFlow Scikit-LearnFlutterIoTREST API

Built real-time CV pipeline for infant emotion detection on custom 5000+ sample dataset — 99.06% train, 86.09% test accuracy with CNN + SVM.
Deployed model on IoT device with REST API backend consumed by a Flutter mobile app controlling 10+ features in real-time.
Published an Indian Patent and industry collaboration with Capgemini for the IoT comfort system design.

~$ ls projects/ --sort=date

Projects

downshift.py ★ new — LLM cost optimization / dev tools

Downshift ↗ github ↗ app ↗ pypi

Sep 2026

Cut LLM costs per PR. Downshift finds every LLM call in a Python repo, tests cheaper models against evals written for each call site, recommends the safe downgrades, and shows the projected cost impact of every pull request.

Built an AST scanner that finds OpenAI and Anthropic call sites and resolves models through constants, imports, env vars and parameter defaults; paired it with an IBM Bob auditor for runtime-only models and shared helpers.
Runs per-call-site evals on the baseline and cheaper candidates and recommends the cheapest model that holds the quality bar; on the demo app, 3 of 8 call sites downgraded for an 18% projected cost cut with no drop in pass rate.
Shipped as a PyPI package and a GitHub Action that posts the projected monthly cost change on every PR; case studies on mem0 and OrchestrAI surfaced 4.5–7.6× more spend than static analysis alone.

PythonASTLLM Evals IBM BobGitHub ActionsNext.js

activation_lens.py ★ flagship — mech interp / AI safety

ActivationLens ↗ github ↗ app

Jul 2026

Mechanistic interpretability & inference-safety research on GPT-2-small — training sparse autoencoders from scratch, building a live harmful-content safety monitor, and optimizing it to run in production.

Trained a sparse autoencoder (768→8,192 features) on GPT-2-small's layer-6 residual stream via TransformerLens; diagnosed 52–82% dead-feature collapse and resolved it to 0.77% via decoder normalization and periodic resampling.
Built a 100-feature regularized probe on SAE activations for harmful-content detection (0.796 AUROC), converted into a live per-token safety monitor achieving 0.758 AUROC and 91.6% recall.
Designed an optimized one-pass monitoring kernel, cutting live-monitor overhead from 4.12% to 0.66% vs. naive full-SAE inference.

PyTorchTransformerLensSparse Autoencoders FastAPINext.js

documentsync_ai.py ★ LLM serving / RAG / Kubernetes

DocumentSync AI ↗ github

Jul 2026 – Oct 2026

A RAG system for document Q&A, self-hosted with vLLM on Google Kubernetes Engine and benchmarked for throughput, latency, and quantization.

Built a RAG pipeline (LangChain, ChromaDB, sentence-transformers) with PDF upload and source-cited answers; RAGAS evals across 4 chunking strategies showed fixed-1024 lifted context precision 59% and recall 3.6× over 256-token chunks.
Self-hosted Qwen2.5-7B with vLLM on an NVIDIA L4: continuous batching scaled throughput 17× (17 to 298 tok/s), prefix caching cut p95 time-to-first-token 67%, and FP8 added 49% throughput with no accuracy loss.
Deployed on GKE with separate CPU and GPU node pools, a LoadBalancer and HPA autoscaling; one OpenAI-compatible client switches between Groq and vLLM through env vars.

vLLMLangChainChromaDB RAGASDockerKubernetesGKE

RAG Eval Sidekick ↗ github

OpenAI Build Week

A full-stack tool built during OpenAI's Build Week to help developers evaluate and debug their RAG pipelines — a Python backend paired with a dedicated frontend for running evals against sample datasets and inspecting retrieval quality.

PythonRAGLLM Evals Full-Stack

Crypto Streaming Pipeline ↗ github

May 2025 – Jun 2025

Real-time pipeline ingesting CoinGecko at 5s intervals. 10K+ records/hour via Kafka + Spark Structured Streaming. 40% latency cut on Athena queries.

KafkaSparkAirflow AWS GlueAthenaS3

Invasive Plant Detection ↗ github

Jan 2025 – May 2025

Transfer-learned industrial defect CNN to invasive species detection — 91.9% accuracy, 2.3× latency reduction via ONNX export (148ms → 64ms on CPU).

PyTorchCNNOptuna ONNXCAM

URL Shortener + Analytics ↗ github

Nov 2025 – Jan 2026

Microservices architecture: 3000+ req/hr at sub-25ms redirect. Redis caching cut latency from 85ms → 20ms. Sub-50ms dashboard queries across 5K+ records.

Node.jsReactRedis MongoDBDockerMicroservices

YouTube ETL Pipeline ↗ github

Jun 2024 – Aug 2024

Serverless ETL on 50K+ YouTube records. Lambda-triggered Glue transforms. 80% less manual prep, 40% query latency cut, QuickSight dashboards.

AWS GlueLambdaAthena S3QuickSight

CodeStreak Tracker ↗ github

Sep 2025 – Dec 2025

Full-stack gamified streak tracker. 100+ active users, sub-200ms response. Auto-verified solves from CodeForces + HackerRank. 90% email delivery reliability.

Node.jsReactMongoDB node-cronREST APIs

~$ cat skills.json