Projects

Real-time wildlife detection and species identification system. Auto-labeled camera-trap imagery with MegaDetector v5a, built a 500-species dataset (~92K train / 8K val images) and fine-tuned a detector to mAP50 ~0.64. Served through a Gradio app that runs on uploaded videos or YouTube links; presented at UB CSE Demo Day.

Multi-tenant SaaS for small businesses: each gets a website, takes customer requests by web or SMS, and sends AI-drafted quotes, invoices and receipts. Includes a lending engine that scores a 6-month bank statement on 8 cash-flow measures in under 300 ms (p95). Load-tested at 2,400+ req/s on one CPU core; runs on AWS for about $12/month with 412 unit/integration and 17 browser tests in CI.

Multi-agent system on IBM watsonx Orchestrate where a supervisor agent routes each question to a specialist: a RAG agent over UB CCR's HPC documentation (56 pages, 96% retrieval accuracy, ~3 s cited answers, embedded on a docs prototype) and a Professor Discovery agent that matches a student's bio to UB faculty research.

3rd Place, IBM SkillsBuild AI Hackathon

One-stop case-management tool for the Against All Oddz animal rescue, built with Team AniTech at UB's AI for Good Design Challenge. Role-based staff and public views, per-animal case pages with a kanban journey, correspondence and urgency indicators; I built request intake across sources.

1st Place, Against All Oddz AI for Good Design Challenge

LSTM, GRU and Transformer forecasters on Dow Jones (DJI) data predicting log returns, with Monte Carlo Dropout uncertainty calibrated to 92% prediction-interval coverage (target 90%), plus Value-at-Risk and Expected Shortfall risk metrics. Presented at UB CSE Demo Day.

Utilizing USA Gov Hourly met data (time-series) to predict Ozone concentration at (t+k) hours using Multi-layer Perceptron with the help of a custom PyTorch Neural Network. Achieved comparable results to baseline time series specialized models like LSTM and GRU

Leveraged Natural Language Processing Concepts to train a vectorizer to successfully classify resume and match to specific job roles. Achieved 83% accuracy.

Built a Convolutional Neural Network classifier to classify skin deformities as malignant or benign, achieving 85% accuracy (comparable to Resnet50 ~86%); reduced false negatives 20%, improving reliability for early detection use cases.

Designed a custom Convolutional Neural Network model trained on manual driving frames from Unity Udacity simulator. The trained CNN model achieved autonomous hands-free driving on the simulator, achieving 92% validation accuracy for steering prediction, with 95%+ lane-keeping accuracy, reduced steering error variance 30% vs. baseline models.

Microsoft Excel VBA code to recognize Barcode, QR Code and Data Matrix output using scanner output and realtime tallying in a large Kit List (~900 parts per model) for Tata Motors South Africa. Active on Industry floor till date. Saved cost-to-company of upto 15000 ZAR

Constructed a complex networking topology with successful end-to-end communication using Routing and Switching protocols