Hi, I'm
I am a Systems Analyst who loves math, finance, and AI. Looking for opportunities at the intersection of it
> Full-Stack RAG System
AI-powered retrieval-augmented generation system for querying SEC EDGAR filings using natural language. Covers the top 10 S&P 500 companies with 1M+ structured XBRL data points and semantic vector search across 10-K and 10-Q filings from 2010 to present.
Intelligent classification routes queries to metric lookup, timeseries, narrative search, hybrid, or full statement retrieval
pgvector cosine similarity search with cross-encoder reranking, combined with structured XBRL financial data
0-100 confidence with 5 weighted signals: retrieval quality, source coverage, cross-source agreement, citations, recency
SSE streaming with live classification display, retrieval plan animation, and progressive answer generation
More projects coming soon.
> Simple Experiment
A simple experiment exploring how poisoned documents in a RAG knowledge base can manipulate multi-agent LLM systems into exfiltrating sensitive data. The simulation uses an orchestrator agent that reads from a shared knowledge base and delegates tasks to a worker agent with database and email access, testing whether injected payloads can trigger unauthorized PII extraction.
Two-agent pipeline with orchestrator and worker roles, mock customer database, and email outbox for measuring exfiltration
Tests multiple attack strategies including direct instruction, authority impersonation, and multi-step social engineering
Chi-squared and Fisher’s exact tests comparing exfiltration rates between poisoned and clean knowledge bases
Measures baseline vs. attacked success rates to quantify how vulnerable current LLM agent architectures are
dev.to · Deep Dive
A detailed walkthrough of every design decision behind the SEC Filing Intelligence Engine — from the naive prototype that hallucinated revenue numbers, through domain-specific nightmares (fiscal year mismatches, XBRL tag renames, Q4 derivation), to the 5-route retrieval architecture that finally worked.
Interested in collaborating or have a question? Reach out.