Everything here is real work, kept in full — a decade of it, from a college code editor to industrial anomaly detection. I've added what each one taught me, because in hindsight almost every project was teaching me the same subject: how to make systems people can rely on.
Applied HBOS, denoising autoencoders, and decision trees to identify operational outliers in streaming data.
What it taught me: detection is the easy half. Every extra false alarm costs you a little of the operator's trust, and trust is the thing the whole system runs on.
Explained individual predictions using SHAP, LIME, and ELI5 for transparent AI decision-making.
What it taught me: if an engineer can't see why a model flagged something, they won't act on it. Explanation isn't a nice-to-have — it's the interface between the model and the decision.
Built classification and regression models — survival analysis, HMMs, RandomForest, XGBoost, RNNs — to predict user churn and performance drops.
What it taught me: predicting decline before it happens is the same math whether the subject is a customer or a gearbox. This was my first contact with the ideas behind remaining-useful-life modeling.
Combined audio signal processing and NLP techniques to auto-generate sports match highlights with 85% precision.
What it taught me: pulling meaningful events out of a long, noisy, continuous signal is exactly the shape of condition monitoring — I just didn't know it yet.
Built a Spray/Undertow-based tool with worker nodes for scalable data ingestion and service orchestration.
What it taught me: a brilliant model behind a pipeline that falls over is worthless. Reliability of the plumbing matters as much as the intelligence flowing through it.
Engineered full NLP pipelines: tokenization, POS tagging, stop-word removal, stemming, lemmatization, intent classification, and entity recognition, plus a minimal utterance generator.
What it taught me: a confidently wrong answer is worse than no answer. Tuning when the bot should say “I don't know” was my first real threshold-setting problem.
Retrained Inception V3 for custom classes and implemented YOLO for real-time object recognition.
What it taught me: “real-time” changes everything. Once inference has a deadline, model choice becomes a systems decision, not a leaderboard decision.
Used Tesseract OCR and KNN clustering on bounding-box coordinates to extract billing details without hard-coded templates.
What it taught me: most real-world data problems are messy-input problems. Skipping the hard-coded templates forced the system to be robust instead of merely accurate on the demo set.
Extracted script regions with OpenCV and classified letters using k-NN with 90%+ accuracy on test datasets.
What it taught me: signal preprocessing decides everything downstream. The classifier was the smallest part of the work; getting a clean signal out of messy ink was the project.
Portal for educators to track attendance, share materials, and automate grading, boosting efficiency by 30%.
What it taught me: automation only counts if people actually use it. The 30% came from watching teachers work, not from clever code.
Developed a mobile social platform for community crime reporting with geotagging and automated alerts.
What it taught me: my first alerting system — and my first encounter with the question that still follows me around: when is an alert helpful, and when is it just noise?
Android application integrating multiple APIs for localized forecasts, push alerts, and interactive maps.
What it taught me: my first taste of turning external signals into timely, localized notifications. Fetch, interpret, decide, notify — the loop I've been building ever since.
Java-based IDE featuring syntax highlighting, auto-completion, and embedded compiler integration.
What it taught me: where it all started. Building a tool other people would sit inside for hours taught me to care about the details nobody thanks you for.
Want the story behind the builder? The full about page has every role, certification, and award.