Serving

vllm-vram-calculator
Tells you whether a model fits on your GPUs, and for how many concurrent users, before you deploy it. Estimates weights, KV cache (GQA/MHA) and activation memory from FP32 down to INT4, AWQ and GPTQ, including tensor-parallel splits. Web UI and REST API.

_{Python · FastAPI · pytest · MIT}

Serving

vllm-load-test
Measures what the deployed server actually does. Runs an async step-sweep load against an OpenAI-compatible vLLM endpoint and records streaming time-to-first-token, end-to-end and inter-token latency, vLLM Prometheus metrics and nvidia-smi GPU telemetry, with per-step summaries and a Markdown report.

_{Python · asyncio · Prometheus · MIT}

Agents & tools

models-mcp
One catalog of a team's model deployments (vLLM LLMs, embedding and reranker models, custom services) that GitHub Copilot can query over MCP. A single FastAPI app serves the MCP endpoint, a REST API and a read-only web catalog. Schema migrations run on startup under a PostgreSQL advisory lock, so replicas can start together.

_{FastAPI · MCP · PostgreSQL · Alembic · Docker · Kubernetes}

Agents & tools

The-Builder
Our entry for the Freya × YC "Build the Builder" hackathon (2026): give it a bank's brief and tool API, and it builds a working Freya phone agent with the Claude Agent SDK. My part: the persona-simulation evaluation harness (LLM and scripted callers, LLM judge) and the final-case test runs. The pass rate went from 6/50 to 27/50 between builds. Built with @resitaydin.

_{Python · Claude Agent SDK · pytest}

Speech + RAG

InsightBridge · _{thesis (PDF)}
My B.Sc. graduation thesis at GTÜ. Whisper transcribes a radiologist's dictation, a domain RAG (ChromaDB + bge-m3) retrieves reference material, and Llama 3.1 via Ollama drafts a structured report in Turkish with findings and impression sections. Includes an evaluation framework.

_{Whisper · ChromaDB · Ollama · LangChain · Gradio · Apache-2.0}

Speech

NarraPhon
Turns a PowerPoint deck's speaker notes (or TXT files, ZIP archives and plain text) into a narrated video with subtitles: StyleTTS2 speech with voice cloning, Whisper subtitles and a Gradio UI. Built on StyleTTS2.

_{PyTorch · StyleTTS2 · Whisper · Gradio · MIT}