ABOUT
Eighteen things I was wrong about.
I build computer vision and retrieval systems, then spend most of the time trying to break them. Four of the claims below were my own conclusions, already written down, killed by the next experiment.
Refuted claims.
| No. | Claim | Refuted by | Source |
|---|---|---|---|
| 01 | An NMS-free detection head is faster to run. | 3.50× faster on TensorRT, 1.69× slower on Core ML | EDGE |
| 02 | That fragmentation is a property of the architecture. | TensorRT compiled the identical graph without issue | EDGE |
| 03 | Latency scales with the number of detections. | 5× more detections cost +0.023 ms | EDGE |
| 04 | Quantization penalises small objects most. | −8.9% small against −9.0% medium on TensorRT | EDGE |
| 05 | Pipeline stages scale with source resolution. | 6.25× the pixels, preprocessing halved | EDGE |
| 06 | A larger embedding model retrieves better. | 110M parameters lost to 22M on every metric | RAG |
| 07 | Insufficient embedder capacity explains the failures. | my own diagnosis, killed by its own follow-up test | RAG |
| 08 | Hybrid retrieval strictly improves on dense. | 2 questions gained, 6 lost at rank 1 | RAG |
| 09 | A similarity threshold can gate hallucination. | the one hallucination scored highest of any question | RAG |
| 10 | The published RRF constant works at any scale. | k=60 returned neither list's top hits at 188 chunks | RAG |
| 11 | Reference noise explains the retrieval failure. | removed 16% of the corpus, ranks 1–4 unchanged | RAG |
| 12 | Claim-level faithfulness scoring catches hallucination. | scored a fabricated comparison 1.0, fully grounded | RAG |
| 13 | End-to-end success means retrieval worked. | correct answer, recall@1 of 0 | RAG |
| 14 | A stride-4 head improves small-object detection. | −0.1477 mAP50, the largest effect measured | VISDRONE |
| 15 | That failure is candidate explosion. | at confidence 0.05, precision rose to 0.553 against 0.463 baseline | VISDRONE |
| 16 | Attention improves feature selectivity here. | uniformly below baseline at every threshold | VISDRONE |
| 17 | Inference-time tiling is universally useful. | +0.0347 on the weakest model, −0.0878 on another | VISDRONE |
| 18 | Class frequency predicts class performance. | 20× fewer instances, 0.27 higher mAP | VISDRONE |
What I work with.
Deep Learning & CVPyTorch, OpenCV, YOLOv8 / YOLO11 / YOLO26, Vision Transformers (Swin, ViT), CNNs (ResNet, EfficientNet), U-Net segmentation, SimCLR self-supervised learning, TensorFlow/Keras, HuggingFace Transformers, scikit-learn
Detection & TrackingObject detection, multi-scale and small-object detection, semantic segmentation, BoT-SORT, ByteTrack, MOTA / MOTP / IDF1 evaluation, SAHI tiling, motmetrics
Optimization & DeploymentONNX Runtime, NVIDIA TensorRT, Apple Core ML, INT8 / FP16 quantization, calibration set design, graph partition analysis, GPU-synchronized latency profiling (p50 / p95)
LLM & RetrievalRetrieval-augmented generation, ChromaDB, BM25, reciprocal rank fusion, cross-encoder reranking, sentence-transformers, Llama 3.1 via Ollama, abstention and faithfulness evaluation, LLM-as-judge
EvaluationmAP@0.5:0.95, ROC-AUC, PR-AUC, macro-F1, mIoU, recall@k, MRR, sensitivity / specificity, controlled ablation design, failure-mode analysis
ProgrammingPython (primary), C++, MATLAB, Java, R, SQL, Bash
EngineeringGit, Docker, FastAPI, Linux, CUDA / MPS, Jupyter, Colab (A100 / T4), pycocotools, Ultralytics, NumPy, Pandas
Education.
M.S. Artificial Intelligence (Biomedical Engineering concentration)
University of North Texas, Denton, TX
January 2024 – December 2025GPA 3.91 / 4.0
B.Tech Computer Science & Engineering
Vijay Rural Engineering College, JNTUH, Nizamabad
August 2019 – May 2023First Class with Distinction
Availability.
RolesOpen to Computer Vision Engineer, Machine Learning Engineer, and Medical Imaging AI roles.
LocationDenton, Texas
RelocationOpen to relocation and to remote work anywhere in the US.
Work authorizationAuthorized to work in the United States through January 2027, STEM-eligible for approximately three years. No sponsorship required now.