Computer Engineering senior passionate about Generative AI, autonomous agents, and RAG architectures β engineering LLM-based applications and workflow orchestrations that solve complex, real-world problems.
- π§ Designing agentic AI systems with multi-agent orchestration (CrewAI) and local LLM inference (Ollama)
- π Building RAG pipelines with vector databases, self-evaluation loops, and hallucination prevention
- βοΈ Fine-tuning transformer models (BERT / DistilBERT) for real-world classification workloads
- π Full-stack capable β React, Flutter, Streamlit, SQL β with a product engineering mindset
- π± Driven to apply AI toward impactful, sustainable global development initiatives
Open To: AI/ML Engineering Internships & Roles Β· LLM Application Development Β· Research Collaborations Β· Open Source Contributions
| Domain | Proficiency | Details |
|---|---|---|
| LLM Applications & RAG | ββββββββββ | LangChain pipelines, ChromaDB vector stores, retrieval orchestration, hallucination prevention with self-evaluation loops |
| Multi-Agent Systems | ββββββββββ | CrewAI agent design β classifier, router, responder architectures powered by local LLMs via Ollama |
| Transformer Fine-Tuning | ββββββββββ | DistilBERT / BERT fine-tuning with HuggingFace Transformers & PyTorch for text classification |
| Deep Learning & CV | ββββββββββ | CNN architectures, transfer learning, multi-class image classification |
| NLP | ββββββββββ | Tokenization, embeddings, semantic search, intent routing |
| Distributed Systems | ββββββββββ | Fault-tolerant event-driven design in Go, Kafka consumer groups, leader election, state recovery |
π« Support Ticket Routing & Multi-Agent Triage
An end-to-end intelligent ticket triage platform combining fine-tuned transformers with autonomous agent orchestration.
| Attribute | Details |
|---|---|
| Stack | DistilBERT Β· HuggingFace Transformers Β· PyTorch Β· CrewAI Β· Ollama |
| Scale | Routes tickets across 10 departments |
| Performance | 71% test accuracy on fine-tuned classification head |
| Security | Fully local LLM inference via Ollama β zero data egress |
| Impact | Automates the classify β route β respond loop end-to-end |
| Repository | Live Demo on Hugging Face Spaces |
A fine-tuned DistilBERT model routes customer-support tickets into 10 departments, connected to a CrewAI multi-agent system with dedicated classifier, router, and responder agents β all powered by a locally hosted LLM. Demonstrates production-style ML deployment from model training to agentic orchestration.
π Ring of the Middle Earth β Distributed Game Engine
A fault-tolerant, real-time distributed system disguised as a two-player strategy game.
| Attribute | Details |
|---|---|
| Stack | Go Β· Apache Kafka Β· Event-Driven Architecture |
| Scale | 3 stateless engine instances in a coordinated consumer group |
| Performance | Real-time gameplay over a Kafka event backbone |
| Security | Stateless engines β no single point of state failure |
| Impact | Full game-state recovery after node termination via event replay |
| Repository | github.com/pxlnstn |
Engineered as a distributed system in Go on an Apache Kafka event backbone: three stateless engine instances form a consumer group with leader election, recovering complete game state from Kafka after any node is killed. A hands-on study in fault tolerance, consensus, and event sourcing.
π AI Evaluation Engine
A local RAG application that audits its own reasoning.
| Attribute | Details |
|---|---|
| Stack | Llama 3.1 (Ollama) Β· ChromaDB Β· Streamlit Β· RAG |
| Scale | Local-first architecture β runs entirely on-device |
| Performance | Self-evaluation loop validates every inference |
| Security | 100% local inference β sensitive CV data never leaves the machine |
| Impact | Actively prevents hallucinations through AI self-critique |
| Repository | github.com/pxlnstn |
A local RAG application that analyzes CV data using Llama 3.1, ChromaDB, and Streamlit β enabling the AI to self-evaluate its inferences and prevent hallucinations. Explores trustworthy AI patterns: grounded retrieval, self-verification, and privacy-preserving local deployment.
π Car Brand Classification
Multi-class visual recognition with CNNs and transfer learning.
| Attribute | Details |
|---|---|
| Stack | Python Β· CNNs Β· Transfer Learning Β· Computer Vision |
| Scale | 33-class car brand classification |
| Performance | Transfer learning for accelerated convergence |
| Security | β |
| Impact | Demonstrates applied deep learning on real-world image data |
| Repository | github.com/pxlnstn |
A deep learning project classifying 33 car brands from images using Convolutional Neural Networks and transfer learning β covering data preprocessing, augmentation, model selection, and evaluation.
learning:
- Advanced Multi-Agent Orchestration Patterns
- LLM Evaluation & Alignment Techniques
- Production-Grade RAG Optimization
building:
- Agentic AI workflows with local LLM inference
- Fine-tuned transformer pipelines for real-world classification
exploring:
- Sustainable AI for global development initiatives
- Distributed systems & event-driven architectures
open_to:
- AI / ML Engineering Internships & Roles
- LLM Application Development
- Open Source Collaboration
- Research Partnerships
