AI Engineer
I design and build scalable, secure end-to-end systems that turn data and AI into measurable business outcomes. My work focuses on architecting reliable, production-grade solutions that solve complex, real-world problems.
Experience
Designing and deploying AI solutions across trading algorithms, customer experience, and operational efficiency.
Scope covers deep learning model development, NLP applications, real-time vector search systems, and generative AI integration — deployed via MLOps tooling and cloud platforms with a focus on model fairness, transparency, and production performance.
First analyst hired; built the company's entire data and reporting function from scratch on a 150+ table SQL Server database with millions of rows.
Designed an executive arrears dashboard covering 3,000+ live cases (CPI, PPI, Trust Deed) with drill-down into arrears, catch-up balances, and assignment-stage pipeline. Built a Node.js API layer to serve query results since direct DB access wasn't permitted for reporting tools, and engineered custom SQL logic with the co-founder/GM to recover ~50 cases silently dropped by the in-house system.
Exposed £60K in previously invisible arrears and catch-up payments, tracked ~£500K in customer payments; dashboard adopted operationally by the co-founder/GM and two business teams.
Replaced rule-based logic with ML inference services for recommendations and classification.
Collaborative (SVD, ALS) + content-based filtering for recommendations; XGBoost + SMOTE for imbalanced classification; wrapped models in FastAPI /predict endpoints.
Engagement +23%, CTR +20%, 92% precision on imbalanced classification in production.
Mentored 120 students in Python and applied ML across guided lab sessions.
Ran labs on preprocessing, visualization, and basic ML algorithms; reviewed code and gave targeted feedback to improve correctness.
Students independently completed data analysis and ML assignments by end of cohort.
Projects
Non-intrusive observer layer for LangGraph multi-agent systems that detects 7 behavioral failure signatures and auto-recovers via checkpoint rollback — 100% task completion across 24 benchmark runs.
ObserverEngine evaluates 7 MAST failure signatures (stagnation, tool loops, context overload, role violations, etc.) over a sliding event window; when a signature persists above threshold, injects a natural-language recovery hint and restores the LangGraph MemorySaver checkpoint to the last stable state. Attaches asynchronously without modifying agent prompts, roles, or tool definitions.
100% task completion across 12 matched observer/baseline pairs; event-count reduction in 7/12 cases (median −14 events); up to 69% token savings on longer runs.
Private AI operating system on a live VPS — 10 autonomous agents each own a domain of personal life or business, controlled entirely through Discord.
10 agents, 3 layers: personal layer (sysadmin, knowledge curator, personal assistant), a dedicated researcher agent for multi-source due diligence, an investor agent with IBKR read access and compliance-aligned screening, and a business C-suite (Chief of Staff, CFO, COO, CTO, CMO). Each calls domain-specific MCP servers backed by PostgreSQL + pgvector; all LLM calls route through LiteLLM + Redis to OpenRouter — models are plug-and-play. Knowledge vault ingests from multiple sources with approval-gated writes. Full observability via Prometheus, Grafana, Loki, and LangFuse.
One VPS, one Discord server — complete personal and business operating layer covering infrastructure, knowledge curation, finance, habits, tasks, research, investment due diligence, and a full business C-suite.
Agentic chatbot that lets users query a MongoDB Atlas database in plain English, combining vector search, full-text search, and document retrieval in a stateful ReAct loop.
LangGraph StateGraph orchestrates an LLM agent over three tools — semantic plot vector search, Atlas full-text title search, and full document retrieval — with MongoDBSaver checkpointing full conversation state and MongoDBStore persisting long-term user memory across sessions. Dual interface: Streamlit web UI and terminal CLI.
Fully stateful multi-turn querying over the sample_mflix dataset; agent autonomously selects the right retrieval strategy per question without user knowledge of query syntax.
Wearable sensor glove that translates ASL hand gestures into speech in real time — 98% classification accuracy with sub-second end-to-end latency.
Flex and IMU sensors embedded in a glove capture finger bend angles and wrist orientation on an ESP32, streaming raw motion data over WebSockets to a cloud-hosted Random Forest classifier trained on labelled gesture sequences. Classified output is passed to a text-to-speech engine and played back in real time, closing the loop from gesture to spoken word without any manual trigger.
98% classification accuracy across the ASL gesture set with sub-second latency from gesture to speech; end-to-end pipeline runs on commodity hardware with no GPU.
Time-series forecasting on 35K hourly records from the Spain national grid — LSTM achieved 1% MAPE, an 83% error reduction over the statistical baseline.
Systematic model comparison across three families on Spain grid data (2015–2018): ARIMA for statistical baseline with differencing and ACF/PACF tuning, Prophet for trend and seasonality decomposition with holiday effects, and a stacked LSTM for sequence-level demand learning. Automated backtesting with rolling windows evaluated each model on identical holdout periods; results surfaced in a Streamlit dashboard with interactive metric overlays and residual analysis.
LSTM achieved 1% MAPE vs. ARIMA baseline — 83% forecasting error reduction; dashboard enables direct side-by-side model comparison across the full backtesting horizon.
Credentials
Comprehensive specialization covering supervised, unsupervised learning, with recommender systems and deep learning fundamentals.
Three-course specialization taught by Andrew Ng covering regression, classification, neural networks, decision trees, recommender systems, and best practices for ML development. Includes hands-on projects implementing algorithms from scratch.
Comprehensive specialization covering linear algebra, calculus, probability, and statistics.
Focused on building the mathematical foundations required for machine learning, including linear algebra, calculus, and probability & statistics. Included practical exercises applying math concepts to real ML problems.
Education
Voluntary Work
Assisted the chair in organizing technical events, workshops, and coordinating teams across initiatives.
Improved event quality and team performance through structured workflows across multiple initiatives.
Led club operations and strategic planning to grow community engagement and streamline event execution.
Improved club reach and participation through collaborative initiatives.
Led technical direction for robotics projects, workshops, and competition preparation.
Executed workshops and achieved strong results at RoboVeda competitions.
Led editorial planning and publication of departmental newsletters (annual and monthly).
Delivered consistent, timely releases under tight deadlines.
Skills
Python 
NumPy 
Pandas Matplotlib Seaborn 
Scikit-learn 
LangChain LangFlow LangGraph 
Django 
FastAPI 
PostgreSQL 
Docker 
Kubernetes CI/CD 
Git 
Bash Agile/Scrum System Design 
PyTorch 
Jupyter 
Hugging Face 
MLFlow 
Streamlit 
Flask 
n8n 
C 
C++