Algorithm Optimization & NLP Toolkit

Engagement Overview

A returning freelance client needed help rescuing multiple algorithmic assignments that were buckling under 10k+ item data files and ad-hoc NLP experiments. I jumped in as the sole engineer, troubleshooting Java heap failures, profiling graph heuristics, and packaging the final deliverables for a presentation-ready submission.

Optimization Workstream

• Reverse engineered the legacy knapsack solver and replaced its dense int DP table with Apache Lucene OpenBitSet buffers, shrinking memory usage from ~37 GB to <6 GB so 10,000×1,000,000 state spaces could run on 64-bit JVMs.
• Added a fallback stream mode that computes the max profit from massive files without materializing every intermediate row, giving the client graded output even on constrained lab machines.
• Hardened file loaders and manifest handling so ks_* and tsp_* datasets could be batch-tested from the provided dist/ folder without manual edits.
• Documented JVM tuning steps (heap flags, GC hints) to reproduce results on the university test rig.

Graph Coloring Enhancements

• Profiled the greedy coloring routine and introduced a highest-degree-first ordering with tie-breaking comparators, which reduced color counts on dense graphs while keeping runtime predictable.
• Added instrumentation hooks so the client could demonstrate optimization strategies and explain algorithm choices during reviews.

Turkish Sentiment Classification

• Reviewed the Zemberek-based morphology idea, then pivoted to a leaner pipeline using unigram frequencies exported from cleaned Excel sheets.
• Built a Weka-backed SVM/Naïve Bayes classifier that ingests positive, negative, and neutral CSV labels and prints predictions for blind tests.
• Supplied scripts for generating top-N feature lists and guidelines for plugging additional LingPipe experiments when more time becomes available.

Impact

• Unblocked every provided dataset, giving the client full marks during the knapsack demo and headroom to discuss optimization heuristics.
• Enabled on-time submission of the combined algorithms + NLP project despite fragmented requirements and limited dataset access.
• Established an ongoing freelance relationship for future research tooling.

Technologies & Concepts