Pretraining & transfer
Self-supervised ViT backbones (DINOv2 → DINOv3) fine-tuned for scientific image-forgery detection; pretrained VGGNet features for captioning; an RNAPro backbone for RNA 3D structure.
AI developer & enthusiast · Phoenix, AZ
Rajesh Sundaram
Computer vision, language models, and self-built agents — turning research into things you can tap, type at, or laugh at.
Ask the AI that knows me
↵
This isn't a search box — it's a little AI that actually knows Rajesh. Ask it anything: his projects, his career, even audit his code — "did he use knowledge distillation?" It runs on his own Hermes agent orchestrator with DeepSeek Pro behind it, grounded in everything public (GitHub, Kaggle, LinkedIn) and cached so most answers cost near-zero tokens.
One secret it's sworn to keep: Rajesh's phone number lives in its knowledge too. Think you can jailbreak it into leaking the digits? Go on — try to break it.
About
Hi, I'm Rajesh — an AI developer and enthusiast based in Phoenix, Arizona. I build language models and computer-vision systems, compete on Kaggle, and run a handful of my own AI agents. I like the part where research turns into something you can actually tap, type at, or laugh at.
Résumé
Twelve years turning machine learning into things that ship — from enterprise software to the research frontier.
~12years in software & ML
1,154GitHub contributions / yr
36Kaggle notebooks
2018published M.S. thesis
Experience & education
-
2018 — now
American Express · Phoenix, AZ
AI & software development inside a large-scale fintech — applying machine learning and modern engineering to real production systems.
-
2016 — 2018
M.S., Machine Learning · Texas A&M University–Commerce
Worked in an ontology lab (knowledge representation & ontology engineering), where my published thesis — deep-learning humour captioning — came to fruition, alongside advisor–student / knowledge-distillation research. Read the thesis ↗
-
2011 — 2015
Enterprise software engineer · SAP ABAP
Began my career building and customising enterprise/ERP software before moving fully into machine learning.
Certifications: Udacity Deep Learning & Machine Learning Nanodegrees · SAP ABAP.
What I actually build with
Not a buzzword list — these are real, from the code in my repos.
Fine-tuning
Fine-tuning large pretrained models with EMA weights and resumable checkpointing; equivariant neural networks (cuequivariance kernels) for RNA geometry.
Curriculum learning
Multi-stage curricula — a 4-stage forgery pipeline and a 6-phase "Singularity" curriculum for ancient-scroll ink detection — with stage-aware validation gating and plateau detection.
Post-training & optimisation
Stochastic Weight Averaging; loss engineering (Tversky · Boundary · clDice · a weighted GlobalCLS master-switch); synthetic data + balanced sampling; homology templates (MMseqs2 + Biopython alignment); advisor–student knowledge distillation — training a small student from a larger advisor.
LLMs & NLP
word2vec embeddings; zero-shot Llama-2 experiments; a perplexity technique for scientific multiple-choice QA; text/humour classification (Naive Bayes, SVM, bidirectional CNN).
Computer vision
U-Net segmentation; noise-residual forgery detection; classical OpenCV contour segmentation; on-device iOS computer vision (Objective-C++).
Agents & tooling
PyTorch / TensorFlow / Keras · polars · Kaggle CLI/API; Cloudflare Pages + Functions; and "Hermes" — my own agent orchestrator running self-healing ops, a WhatsApp assistant, and continuous security monitoring of my own systems and dependencies.