Machine learning and computer vision, built into real products

What I have built

• Face recognition. FaceAttend uses InsightFace ArcFace embeddings (512 dimensions) with YuNet face detection, multiple embeddings per student and a configurable similarity threshold. Read how it works.
• Audio fingerprinting. SoundFind builds spectrogram peak constellations, hashes peak pairs and matches them against your own catalogue in a FastAPI service. Read how audio fingerprinting works.
• Reinforcement learning. SecureOps uses reinforcement-learning firewall decisions with a retraining threshold on Azure ML.
• Parallel computing research. A replication of the OMP4Py paper with FFT, Jacobi, LU decomposition and molecular dynamics benchmarks.

How I work on ML projects

1. Define the decision the model supports and how you will know it is right.
2. Build the simplest pipeline that works end to end, then improve accuracy.
3. Put monitoring and retraining in place before calling it done.

What you get

• A clear definition of the decision the model supports and how accuracy will be judged.
• An end-to-end pipeline first, then accuracy improvements measured on held-out data.
• A threshold tuned to the real cost of false matches and missed ones.
• The model wrapped in a service your product can call, with monitoring around it.
• A plan for retraining or re-enrolment as data changes.