REZINX
Transparent AI Solutions
Interactive Portfolio
Real-world AI systems built for production. Explore detailed case studies, live telemetry simulations, technical constraints, and measurable impact metrics.
AI Engineer & founder
AI engineer with three years of experience building production AI systems: LLM fine-tuning, RAG pipelines, and full-stack ML backends with complex optimization layers. I've worked as the sole engineer on end-to-end projects for international clients, handling everything from raw data to deployed services. Currently completing a B.Sc. in Artificial Intelligence.
"I care about the parts that don't fit on a slide: handling sparse and messy data, cutting inference latency, indexing high-dimensional vectors, and shipping services that hold up under real load, not just notebooks that run once."
Our flagship Final Year Project. Features real-time human pose-based fall detection, hazard recognition, child-safety monitoring, and a RAG-powered first-aid guidance assistant (powered by FAISS & BM25 vector search) hooked to Firebase FCM notifications.
We build taxonomy-aware retrieval: embeddings organized around semantic hierarchies instead of raw objects, on a re-architected Qdrant layer. Typical result is a 3–5× latency reduction on large catalogs.
A production price-elasticity optimization engine. Ingests catalog logs into Supabase, performs SKU demand forecasting using LightGBM, and passes predictions to a two-phase LP + SLSQP profit optimizer under margin floors and competitor guardrails. Includes champion/challenger WMAPE gating.
Custom model deployments for specialized clients. Led LoRA and QLoRA fine-tuning workflows for custom domain-specific generative models, automated model training pipelines on DigitalOcean droplets and physical GPU nodes, and shipped custom natural language and vision pipelines.
Engineering Case Studies
Not a list of job duties. Each item is a real engineering bottleneck we hit, what we built to resolve it, and the performance output it generated.
Embedding around semantic taxonomies instead of raw objects, plus a re-architected Qdrant layer.
Async request handling lifted a hard concurrency ceiling without rewriting the models.
The optimizer doesn't scan every price. A two-phase LP + SLSQP search explores a bounded band around each SKU's current price (shaded), under margin floors and competitor caps, to maximize profit, not revenue. Demand falls as price rises; profit peaks where volume and margin trade off, and turns negative below unit cost.