Gen-Ai-practice-coder-Army-

A collection of hands-on Gen-AI / generative-modeling projects — every one runs standalone on a laptop CPU.

Projects

Project	Main file	What it is
flow-matching-toy	train.py	Toy Flow Matching implementation — learning vector field from noise to circle distribution (CERN GENIE related). The core math behind diffusion/flow models like SD3.
llm-hallucination-judge	eval.py	Lightweight LLM evaluation framework — hallucination detection without API calls. Pure-Python claim-level faithfulness judge.
pytorch-vae-mnist	vae.py	Minimal Variational Autoencoder on MNIST — reparameterization trick, ELBO loss, latent-space visualization.
calo-shower-cfm	src/train.py	Project 4 — Conditional Flow Matching for calorimeter shower generation: rectified-flow ODE dx/dt = v_θ(x,t,c) conditioned on incident energy/direction, trained on synthetic 3-layer ATLAS-style showers, samples real-vs-generated images.

How to run

Each project folder has its own README with instructions:

# Project 1 — needs torch + matplotlib
cd flow-matching-toy && python train.py

# Project 2 — needs NOTHING (pure stdlib)
cd llm-hallucination-judge && python eval.py

# Project 4 — needs torch + matplotlib
cd calo-shower-cfm && python -m src.train --steps 2000   # training
cd calo-shower-cfm && python -m src.sample --n 4         # sampling + plot

Topics covered

Flow matching / conditional flow matching • ODE-based generative sampling • hallucination evaluation & faithfulness judging • claim extraction • VAEs • reparameterization trick • ELBO / KL divergence • latent spaces • calorimeter shower generation (project 4)

Tested & verified ✔

All four programs were run end-to-end (no API calls for project 2, CPU-only for 1, 3 and 4) — once in an integrated PowerShell terminal and once inside VS Code using the pre-configured tasks in .vscode/tasks.json. Eight runs, eight exit code 0. Full gallery with per-run descriptions: screenshots/

Project	VS Code terminal run	Result
flow-matching-toy		4000 steps, final loss: 1.02470
llm-hallucination-judge		3/5 hallucinations flagged, 0 API calls
pytorch-vae-mnist		ELBO improved -165.51 → -157.44, plots saved
calo-shower-cfm		9/9 CFM tests passed, loss 1.01 → 0.018, generated showers saved