Downshift  ·  live demo  ·  PyPI Cuts LLM costs per PR. Finds every OpenAI and Anthropic call in a Python repo via AST analysis, tests cheaper models against evals written for each call site (181 cases), 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. On mem0 and OrchestrAI, the agent audit surfaced up to 7.6x more projected spend than static analysis alone.

Python AST Analysis LLM Evals GitHub Actions Ollama IBM Bob

ActivationLens Mechanistic interpretability on GPT-2-small: trained a sparse autoencoder on the layer-6 residual stream, cut dead-feature collapse from 52-82% to 0.77%, then built a live per-token safety monitor (0.758 AUROC, 91.6% recall) with a one-pass kernel that trimmed monitoring overhead 6x (4.12% to 0.66%).

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 LLM-as-Judge

DocumentSync AI RAG Q&A with Qwen2.5-7B self-hosted on vLLM on GKE (NVIDIA L4). Continuous batching scaled throughput 17x, FP8 cut p95 latency 38% with no accuracy loss, and RAGAS evals across 4 chunking strategies lifted context precision 59% and recall 3.6x.

vLLM Kubernetes Docker GCP LangChain ChromaDB RAGAS

Real-Time Crypto Streaming Pipeline Streams live market data via Kafka + Spark Structured Streaming into a Redshift star schema, orchestrated with Airflow. Ran 72 hours continuously landing ~468K records, with dead-letter routing and alerts within 5 minutes; cut Athena bytes scanned by 89%.

Kafka Spark Airflow AWS Redshift