u/adnanqidwai/projects.md
projects
A few projects I have worked on, grouped roughly by area.
AI Systems, RL and Agents
- PatchGym
Verifier-backed coding-agent environment that generates seeded Python repair tasks, runs agents in copied workspaces, and grades patches with public/hidden tests, API checks, and edit budgets.
- RLM + GEPA for Retrieval
Retrieval-policy harness where a DSPy RLM searches local Markdown documents, cites evidence, and is scored from tool traces. GEPA refines the written policy from answer, citation, and budget failures.
- TinyRLVR
Small verifiable reasoning lab for Countdown/24-game tasks, exact arithmetic checking, expression-tree policies, and a Q-learning controller for budgeted test-time compute.
- LLM Compiler
LangGraph tool-use agent that decomposes a query into dependent tool calls, runs independent calls in parallel, and joins the results before answering.
- Voice RAG Tutor
Voice-enabled learning tool that combines retrieval, agent routing, and text-to-speech for learning about music.
Compilers and systems
- Nanopass Compiler for Racket
Racket-to-x86 compiler organized as small lowering passes across staged IRs, with vectors/tuples, heap allocation, garbage collection, and register allocation.
- Parallel t-Spanner Construction
C++ graph-algorithms project for constructing sparse t-spanners and comparing sequential versus parallel construction behavior.
- Eggshell
Unix-style shell implementation with command execution and systems-programming primitives.
Core NLP Tasks
- Neural POS Taggers
Part-of-speech taggers built with feed-forward and recurrent neural models for sequence-labeling experiments.
- NLP Course Exercises
Assignments from Stanford 224n covering word vectors, neural classifiers, and sequence models.
Classical ML
- K-Means from Scratch
K-means clustering from scratch on a football dataset, including centroid updates and cluster assignments.
- Multiclass SVMs
Multiclass SVMs comparing one-vs-rest and one-vs-one classification strategies.
- Random Forest from Scratch
Random forest classifier from scratch over penguin data, covering bootstrapping, feature splits, and tree-vote aggregation.
- Clustering Algorithm Comparison
Comparison of clustering algorithms across compact, skewed, subclustered, and well-separated datasets.
Data mining and analytics
- Paged Data-Cube Construction
Paged data-cube construction with merging and attribute-oriented induction.
- FP-Growth Movie Recommender
Movie recommender built with FP-Growth association-rule mining over frequent itemsets.
- Makaan Housing Analytics
Housing-data analysis over scraped Makaan.com listings, covering cleaning, summary statistics, and basic market patterns.