Syam Sundar Nallamekala

Research

I publish what survives honest evaluation, including when that means retracting a result.

Three published papers and one granted patent. Two of the papers are hybrid architectures applied to imaging where the boundary is genuinely gradual; the third is the undergraduate robotics work the patent came from. What follows is a thesis in progress, and an audit that invalidated a prior result before it invalidated itself.

Publications

What I have published

  • Neural Computing and Applications, Springer Nature

    Fuzzy deep LSTM for cancer cell death analysis

    Published — citation pending

  • MS thesis research, University of Toledo

    Hip osteoporosis risk from CT imaging and clinical data

    Published — citation pending

  • 2022 6th International Conference on Electronics, Communication and Aerospace Technology (ICECA), IEEE, pp. 1153–1157

    FlowerBot: a deep learning aided robotic process to detect and pluck flowers

    Published 1 December 2022 — cited by 7

  • University of Pennsylvania

    External validation

    The brain tumor classifier was validated externally, outside the group that built it.

Patent

Flower Plucking Robotic Arm

The earliest work on this site, and the one that predicts the rest of it: vision applied to a messy physical process where the measurement is the hard part. Rose picking had not been automated — thorn injuries, sustained bending, and no visibility after dark.

Patent
IN 573184
Application
202241055456
Filed
28 September 2022
Granted
31 October 2025
Patentee
Kalasalingam Academy of Research and Education
Inventor
Nallamekala Syam Sundar
Extent
11 pages, 6 claims

The project on Projects

Poster

Fuzzy deep neural network for brain tumor detection and classification

University of Toledo, Dept. of EECS — with Dr. Devinder Kaur

Research poster: fuzzy deep neural network for brain tumor detection and classification, showing the methodology, ROC curves and per-class performance metrics.
The hybrid reaches an ROC AUC of 0.926 against 0.871 for a CNN alone and 0.815 for traditional ML — the comparison matters more than the headline accuracy, because it measures the hybrid against what it replaces.

How I work

Method, and what I do when it fails

  • I audit inherited work before I extend it

    The osteoporosis pipeline came to me from a prior student. Auditing it end to end surfaced a data-leakage bug — the model was effectively seeing test data during training. After correction a previously reported headline result did not survive, and a simpler model proved the more trustworthy one.

  • I replace single splits with repeated cross-validation

    One train/test split made the initial osteoporosis results look strong. Repeated cross-validation showed most of that was noise, and confirmed genuine signal only under stricter testing.

  • I report conditions alongside numbers

    Every figure on this site carries what it was measured on. The traffic model reports 8.92% MAPE and, in the same breath, that it has 38,375,647 parameters trained on six examples and no baseline to beat.

  • 480+ documented research hours at ODOT

    Across three assignments, with three milestone presentations to stakeholders and a final technical report alongside the delivered prototype.