I build AI systems that ship to production.

Ciel runs realistic, multi-round mock interviews with a committee of AI interviewer agents. You search a live role or paste a job description, and it builds the interview from the company's real hiring process, scores you as you go by voice, text, or code, then turns your weak spots into a study plan. Above the committee sits a hyperagent, a meta-agent that improves the interviewers' own questioning and scoring strategy across generations, and even rewrites how it improves.

A peek inside

The committee

By the numbers

How it's built

The hard part. The moat is the self-improvement loop: a hyperagent that mutates and re-scores its own interviewers, with a train/validation split and a held-out discrimination benchmark it is never allowed to see, so gains are real and not self-flattering.

Typed monorepo, a large automated test suite, and CI-style typecheck, test, and build gates on every change. Closed-source; a guided code walkthrough is available on request.

A command-line tool I designed, built, and maintain. It inspects a Claude Code context window, identifies where tokens are being wasted, and generates actionable fix commands. It runs fully offline with zero API calls, which keeps it fast and private. Published to npm and adopted by developers in the wild.

LLM-powered analysis system for NVMe semiconductor test data

• Architected an LLM analysis system using fine-tuning (LoRA/QLoRA) over millions of log entries to automate narrative generation, reaching 85%+ narrative quality and cutting manual analysis from hours to ~6 minutes.
• Built a 3-layer hybrid retrieval architecture combining SQL queries, BM25 sparse retrieval, and dense semantic search with Reciprocal Rank Fusion (RRF).
• Engineered an LLM guardrails and observability layer with hallucination detection and automated quality scoring, holding numerical accuracy above 95%.

ML-driven dashboard for error forecasting and anomaly detection

• Built error forecasting, anomaly detection, and root-cause insights for testers.
• Created an ARIMA-based model with Python UDFs in Snowpark ML for time-series forecasting at 84% accuracy.
• Enhanced anomaly detection with autoencoder-based UDFs to flag outliers across 2.5M+ daily test entries, with integrated root-cause diagnostics.

Internal tester dashboard, full-stack (Flask + D3.js)

• Resolved a browser crash from a 200MB+ bulk load, cutting page load from 20+ minutes to under 10 seconds by redesigning the Snowflake query architecture.
• Built a FastAPI + Streamlit clustering tool with deep-linking, auto-zoom, CI/CD, and drill-down modals, backed by a 27-test suite, used by teams to pre-screen issues before debugging.

Driver drowsiness detection

• Built an enhanced drowsiness detection system with a hybrid metric fusing Eye Aspect Ratio (EAR), PERCLOS, and Gaze score; used GANs for face synthesis and GPT-3 to refine feature extraction, for an 18% accuracy gain.

Multi-step vehicle trajectory prediction (radar)

• Designed an algorithm fusing an Unscented Kalman Filter (UKF) with a Spatio-Temporal Graph Convolutional Network (ST-GCN) to predict on-road vehicle trajectories from radar sensor data.

Web-scraping and data-extraction tools

• Built a Google Patent parsing tool with the BeautifulSoup Python package to extract structured data from HTML.
• Used Selenium WebDriver to scrape sites with dynamic, JavaScript-rendered content.
• Implemented text-inconsistency detection software.

TraceX — NLP conversational interface for COVID-19 risk

• Led the design and build of TraceX, an NLP conversational interface integrated with Telegram (400M+ user base) for real-time COVID-19 risk assessments through interactive dialogue.
• Implemented conversational flows in GCP Dialogflow with intent structuring and entity extraction for personalized risk evaluations.
• Fine-tuned an ML information-retrieval model over news articles (precision 0.78, AUC 0.80) to predict infection risk.

Cyberbullying detection (deep learning)

• Designed and trained a hybrid RNN-LSTM network, using BERT for data augmentation and GPT-2 for synthetic data generation, improving accuracy by 2.4%.

Led administration, budgets, and HR operations across a 17-person cross-functional team in an Agile environment, overseeing departmental resource allocation.