Software Engineer · Python · Generative AI · Machine Learning
I'm a software engineer at Nuvogen, working across generative AI, machine learning and backend engineering in Python.
Most of my work has been on SaaS products, including ReviveMyReputation (formerly OnlineReputation) and Edge-Forex, where I led development. Before that, I worked as a software engineer with Edge-Forex's trading team.
Outside work I build projects from first principles in information retrieval, computer vision and systems programming, from a hybrid search engine to a memory allocator. I also wrote a technical report on data leakage and evaluation in industrial defect detection.
Away from the keyboard, I read about international relations and global politics, in publications such as Foreign Affairs, Chatham House and the Council on Foreign Relations. I'm also drawn to human psychology, decision-making and how societies think and behave, and enjoy the essays in Aeon and Behavioral Scientist.
Independent work, done outside my job. I'm interested in retrieval, generative AI and how ML systems are evaluated, especially when a benchmark number doesn't match how a system behaves in practice.
The official split of a public casting-inspection dataset leaks 71.7% of its test images into training. I rebuild a clean split, train a small CPU-friendly baseline, and look carefully at what its test results can and cannot show.
Dense + sparse retrieval combining FAISS, BM25, RRF fusion, and cross-encoder reranking. Built to handle both semantic similarity and exact-match queries that pure dense retrieval misses.
An in-memory, multi-threaded ledger used to study lock granularity, deadlock-free lock ordering, and CAS versus locking under uniform and hotspot contention, with JMH benchmarks and stress tests that check the invariants.
Chunking decides what a retriever can find and what a generator can see. A look at fixed, recursive, semantic and structure-aware chunking, the precision/recall trade-off behind chunk size, parent-child and contextual retrieval, and how to evaluate a chunking configuration instead of guessing.
A walk through the full path of a decoder-only transformer, from bytes to tokens to embeddings, through RoPE, attention, feed-forward layers and residual streams, to the training objective, sampling, and the arithmetic that sets the cost of training and inference.
Why a small model trained on a large model's outputs learns more than one trained on the data alone. Soft targets and temperature, the distillation loss and its T² factor, forward versus reverse KL for generative models, feature-based and on-policy distillation, and how distillation combines with pruning and quantization in production.
The KV cache turns quadratic recomputation into linear work, then becomes the main memory cost of serving LLMs. Where K and V come from, prefill versus decode, how to size the cache, and the techniques that shrink or manage it, from GQA and MLA to PagedAttention, prefix caching, quantization and eviction.
Why a model trained to predict text is not a model that does what we want, and how post-training closes part of that gap. Supervised fine-tuning, preference data and the Bradley–Terry reward model, RLHF with PPO and a KL penalty, the derivation of DPO, GRPO and verifiable rewards, AI feedback, and the failure modes such as reward hacking and sycophancy that remain open.
Leading development of SaaS products including ReviveMyReputation and Edge-Forex. Previously a software engineer with Edge-Forex's trading team.
Worked across the backend and ML sides of the product, shipping features end to end. Converted to a full-time offer after the internship.
