Software Engineer, Applied AI at Cognida
I take software from the first customer conversation to production, bringing together applied AI, product thinking and full-stack engineering.
I build AI products end to end — the model pipelines, the services around them, and the interface users touch. My work lives where a convincing demo has to become a real product: ambiguous requirements, messy data, and systems that hold up beyond the happy path.
As a startup engineer I own the full arc of a feature: understanding the problem, shaping the product, then building it across backend, frontend, integrations and LLM pipeline. Before this I worked on more traditional systems, turning complex problems into reliable software. That grounding is why the software, data and product decisions around the model get as much of my attention as the model itself.
What interests me most are the decisions that turn an AI capability into a useful product — what the model should handle, what belongs in conventional software, and how much complexity the problem warrants. I look for the simplest thing that solves the real problem, then make it work in practice.
Architected a persistent context graph for enterprise AI agents, synthesizing data across many sources into interconnected, source-grounded context, with retrieval that combines vector search and graph traversal. Engineer reliable LLM pipelines with schema-constrained outputs, deterministic validation and confidence-based escalation to human review. Build AI-native automation for finance and legal operations, owning delivery end to end across Django services, integrations and frontend.
Designed and shipped a Bulk User Management feature for Amazon Ads, backed by scalable Spring Boot APIs, improving account operations across 10K+ advertiser accounts.
Built an interactive ReactJS interface with AI-generated visualizations for analyzing human behavior in video, supporting clinical diagnostic research.
Developed a gamified quiz module on a real-time multiplayer learning platform serving 1000+ concurrent users, lifting engagement by 25%.
Built a computer vision model in PyTorch to identify product label defects in real time on packaging lines, running inference on edge-captured images.
Built an inventory management dashboard for the manufacturing supply chain — responsive Angular frontend, Java Spring Boot APIs, PostgreSQL — with an LSTM time series model for demand forecasting.
Developed machine learning models for supply chain prediction, working with historical inventory and logistics data to forecast demand.
Master of Computer Science
Bachelor of Technology, Computer Science Engineering
An LLM-based writing assistant that mimics a user’s writing style. Fine-tunes open-source Llama and Mistral models with LoRA and layers a RAG pipeline at inference. Quantization keeps the whole pipeline trainable locally on your own data.
View repository →Generating statistically accurate synthetic images for domain-specific applications in medicine and material science. Extends StyleGAN2-ADA with new regularization losses that actively learn domain-relevant statistics.
View repository →Automated ball action spotting in soccer matches, localizing and classifying passes, shots and goals with high temporal precision. A hybrid CNN + Transformer architecture trained on the SoccerNet dataset.
View repository →A tool for analyzing human behavior in video for clinical diagnostic applications, using GPT-4o for multimodal video and language understanding.
A forecasting model for store sales that accounts for seasonality and trend, drawing on sales history, holidays, oil prices and transactions. Combines hypothesis testing and feature engineering with both statistical and ML models.
View repository →A computer vision smart-home system for locating household items, built on a YOLO-v4 detector trained on a custom dataset and wrapped in a voice-enabled web app for accessibility.
Applying Differential Evolution with the NSGA-II algorithm to multi-objective optimization, searching neural network architectures against competing objectives rather than a single score.
View repository →Exploring listwise reranking methods that use large language models to reorder retrieved documents by relevance, improving the quality of retrieval pipelines.
View repository →Network analysis of the arXiv high-energy physics theory citation dataset — graph structure, centrality and community detection across a large corpus of research papers.
View repository →If you're building something where AI has to hold up in front of real users, I'd like to hear about it — roles, interesting problems, or just a conversation. The fastest way to reach me is email.