Incoming Founding AI & Data Engineer

The Challenge

School districts lacked predictive visibility into IEP compliance, academic progress, and operational risk — relying on slow, fragmented manual data aggregation across disparate assessment platforms.

Architecture

• Designed a LangGraph agentic pipeline orchestrator with gate-node conditional routing, fan-out parallelism across 4 concurrent stages, and a local Ollama LLM (qwen2.5:7b) that analyzes failures and suggests fixes — no student data leaves the server
• Built a 6-domain ML prediction system (CAASPP ELA/Math, ELPAC, Chronic Absenteeism, College/Career, Suspension) using scikit-learn, ARIMA, and Holt-Winters with per-district model selection; integrated OpenAI + Gemini APIs for plain-language administrative narratives
• Automated IEP compliance tracking across 21 districts: Playwright PDF download with MFA handling → PDF date parsing → green/yellow/orange/red deadline risk scoring → Power BI dashboards; zero missed IEP deadlines after deployment
• Engineered 32 production tools covering ETL, compliance reporting, browser automation (Playwright + Selenium), async web scraping, SFTP delivery, a Go REST API client to MongoDB, and 150+ Power BI dashboards across 9 dashboard types
• Built automated Gmail API email alert system delivering data-driven weekly performance summaries to 18 school sites with dynamic metric selection from 12 indicators

Business Impact

Platform deployed across 30+ California school districts. LangGraph agentic automation achieved a 75–90% reduction in pipeline processing time. Replaced dozens of hours of weekly manual work — data collection, compliance tracking, report generation, and stakeholder communication — with single-command automated pipelines.

The Challenge

Transparent proxies silently intercept and modify web traffic without user awareness, but their true prevalence, behavior, and network-wide impact were poorly understood at scale.

Methodology

• Designed a large-scale active measurement system to detect and fingerprint transparent proxies across global internet paths
• Built Python-based data collection and analysis pipelines processing millions of network probes
• Developed novel detection heuristics combining HTTP header analysis and TCP-level signals

Academic Impact

Published at IEEE INFOCOM 2024 — one of the top-ranked venues in computer networking, revealing the significant hidden influence of transparent proxies on internet traffic integrity.

The Challenge

The open proxy landscape — used for anonymization, censorship circumvention, and malicious activity — had never been comprehensively characterized in terms of scale, geography, and behavior.

Methodology

• Crawled, scanned, and analyzed 436,000+ open proxies across the global internet
• Built large-scale data collection infrastructure using Python and AWS for distributed scanning
• Applied statistical modeling and traffic analysis to characterize proxy behavior, uptime, and abuse patterns

Academic Impact

Published in Computer Networks (Elsevier), 2022 — delivering the first comprehensive analysis of the open proxy ecosystem and its security implications at internet scale.

The Challenge

Remote peering in BGP networks was known to distort anycast routing decisions, but the extent of this unintended impact on global traffic distribution — including for major cloud providers — had not been passively quantified.

Methodology

• Developed a passive BGP measurement methodology to infer anycast catchment boundaries without active probing
• Analyzed global BGP routing tables and AS-path data across hundreds of vantage points
• Correlated routing anomalies with remote peering relationships at internet exchange points (IXPs)

Academic Impact

Published in ACM SIGCOMM Computer Communication Review, 2019 — a flagship networking venue — establishing foundational methodology for passive anycast analysis used in subsequent internet measurement research.

The Challenge

As AI crawlers become ubiquitous, websites are moving beyond binary blocking (robots.txt) to a more sophisticated, unmeasured tactic: returning HTTP 200 responses to both humans and AI bots, but serving degraded, watermarked, or "poisoned" content specifically to crawlers like GPTBot.

Methodology

• Twin-crawler framework (Playwright) visiting Tranco Top 10,000 domains — once as a standard browser UA, once as GPTBot
• DOM tree structural comparison and text similarity scoring via Jaccard & TF-IDF cosine similarity
• Sector-level taxonomy: paywall injection, text truncation, gibberish poisoning, visual watermarking

Novelty

Unlike prior work measuring blocking, this measures deception — filling a critical gap in understanding how the web's content landscape diverges between human and AI readers.

The Challenge

LLMs are widely used to generate Infrastructure-as-Code (Terraform, Kubernetes YAML, Nginx configs). If a model hallucinates a plausible but unregistered domain endpoint, an attacker could register that domain to intercept live API traffic or credentials from deployed systems.

Methodology

• 1,000+ DevOps-focused prompts submitted to GPT-4o, Claude 3.5 Sonnet, and Llama-3-70B
• Regex extraction of all generated domains, filtered against known public registries
• DNS resolution + Registrar API queries to quantify hallucination rate and live registrability of phantom endpoints

Novelty

Distinct from package hallucination studies — this targets DNS-level infrastructure interception, a critical supply chain risk not previously measured in the LLM security literature.