CSE-Research

Two research artifacts from the compilers and program-analysis group at IIT Kanpur, each with a deployed application built on top of it.

Live applications

→ cyclops-delta.vercel.app

Cyclops — LL(1) grammar analysis and parse-table workbench. Enter a grammar, get FIRST sets, FOLLOW sets and the parse table, then fill the table in yourself and have it graded cell by cell with the governing rule named for every mistake. Next.js on Vercel.

→ nnrepair.streamlit.app

NNRepair — constraint-based repair of neural network classifiers. Browse the 345 published result files, decode the Z3 solutions into weight deltas, and re-run the repaired networks over the complete datasets. Streamlit Community Cloud.

NNRepair deploys from shashvat-singham/nnrepair-app, a split-out copy of apps/nnrepair. That keeps the build off this repository's ~1 GB of raw research data while still shipping every artifact the app needs — the input datasets travel losslessly compressed, 941 MB down to 29 MB.

Layout

apps/
  cyclops/        Next.js app — the LL(1) engine and workbench
  nnrepair/       Streamlit app — results explorer + the Python port
Cyclops/          original artifact (CRA + Express + PHP + Python)
NNRepair/         original artifact (Java, constraints, Z3 solutions, results)

The originals are left in place. The apps under apps/ are what deploys.

Why two different hosts

They are different kinds of thing, and forcing both onto one host would make one of them worse.

Cyclops is an interactive app: routing, an editable parse-table grid, per-cell validation. That is a web front end, and Streamlit cannot express the grid.

NNRepair is data exploration over 345 result CSVs plus numerical code someone might want to actually run. That is Streamlit's core competence, and it tolerates compute and data sizes that Vercel's serverless limits do not.

Getting started

Docker

Both apps together:

docker compose up --build

	URL
Cyclops	http://localhost:3000
NNRepair	http://localhost:8501

Or one at a time:

docker build -t cyclops ./apps/cyclops   && docker run --rm -p 3000:3000 cyclops
docker build -t nnrepair ./apps/nnrepair && docker run --rm -p 8501:8501 nnrepair

Both images are multi-stage, run as a non-root user, and carry a healthcheck that exercises the app rather than just the port. Cyclops builds Next's standalone output, so the runtime image holds only the modules actually imported; NNRepair installs into a virtualenv in the builder stage and copies it forward, leaving no compilers or pip cache behind.

Compose bind-mounts NNRepair/Z3Solutions and NNRepair/NN-Code read-only, which lights up the two NNRepair pages that need them. They are far too large to bake into an image, so docker run without the mounts gets an app that explains their absence instead.

Without Docker

# Cyclops
cd apps/cyclops && npm ci && npm run dev

# NNRepair
cd apps/nnrepair && pip install -r requirements.txt && streamlit run streamlit_app.py

Repository size

.git is ~457 MB and the working tree ~1 GB, almost entirely NNRepair/NN-Code: ten MNIST/CIFAR dataset files of 94 MB each, plus extracted weight dumps. Cyclops/node_modules was also committed — 42,237 files.

Cyclops/node_modules has been untracked — it is reinstallable and never belonged in version control. NNRepair/NN-Code is still tracked: it is research data, and dropping it would deny it to anyone cloning the repository. .gitignore now covers both, so neither grows further.

Shrinking .git itself means rewriting history with git filter-repo, which changes every commit hash and needs a force-push. Worth doing, but that is a call for whoever owns the remote — not a side effect of this work.

Neither deployment includes the large data. See apps/nnrepair/README.md for which pages need a local checkout.

Notes on the ports

Both apps rewrote code rather than wrapping it, and both READMEs record what changed and why:

• Cyclops — the old server ignored its own arguments and returned one hardcoded grammar for every request; the parse-table generator loaded grammars from an absolute path on a machine that no longer exists; FIRST/FOLLOW recursed with no base case under a recursion limit of 80.
• NNRepair — the Java is ported to NumPy and verified against literal transcriptions of its own loops to 1.6e-14. It also documents a ~1.5 point discrepancy between re-running the pipeline and the shipped result CSVs, which traces to the committed weights/datasets rather than to the port.