Applied AI & Systems Engineer • Building measured, verifiable AI systems and developer infrastructure.
LinkedIn • Portfolio • X (Twitter) • GitHub
I design and build autonomous agent pipelines, document intelligence engines, and developer infrastructure focused on reliability, deterministic verification, and measurable performance.
- Deterministic Verification over Blind Trust: Pairing probabilistic LLM generation with sandboxed formal validators (Python AST execution, Luhn check algorithms, clinical taxonomy constraints) to eliminate mathematical and logical hallucinations.
- Empirical Benchmarks over Intuition: Evaluating retrieval and agent architectures against established industry benchmarks (FinanceBench, FHIR R4 conformance, RRF ablation studies).
- Zero-Bloat Architecture: Prioritizing in-memory algorithms, pure Python AST traversal, and lightweight data structures before introducing heavy distributed database daemons.
- Mintplex-Labs/anything-llm (35k+ ⭐)
Audit-Grade Financial RAG Engine & Deterministic SEC 10-K Analysis Terminal
- Eliminates financial hallucinations across corporate filings by pairing table-aware structural chunking with a deterministic Python AST math engine (
decimal.Decimal). - Live Ingestion: Directly streams and parses real Form 10-Ks from SEC EDGAR over HTTPS with fair-access rate limiting and content-addressable disk caching.
- Empirical Results: Increases Answer Accuracy from 5.0% to 90.0% and Evidence Recall@1 to 90.0% across 20 FinanceBench corporate disclosures.
- Stack: Python, FastAPI, NumPy, Rank-BM25, SEC EDGAR API, Python AST, Vanilla JS/CSS.
Clinical Document IDP Engine & Calibrated Clinician Review Terminal
- Production-grade Intelligent Document Processing (IDP) pipeline for ambulatory medical referrals and facsimile forms with distribution-free conformal risk control.
- Deterministic Clinical Invariants: Enforces closed-form verification—CMS 10-digit NPI Luhn check digit formula, CMS ICD-10-CM active taxonomy registry, and timeline chronology.
- Pluggable Architecture: Works 100% offline with zero API keys via native hardware-accelerated Windows OCR / Tesseract, or seamlessly plugs into Gemini 2.5 Flash, GPT-4o, or local Ollama.
- Interoperability: Auto-accepts verified records into standard HL7 FHIR R4 US Core bundles; routes ambiguous fields to a split-pane clinician adjudication workstation with interactive SVG bounding boxes.
- Stack: Python, FastAPI, Pydantic v2, Windows Native OCR, Conformal Risk Routing, HL7 FHIR R4, Docker.
Deterministic Fault-Injection Harness & Invariant Verifier for LLM Agents
- Eliminates duplicate customer refunds, phantom return labels, and race conditions in autonomous agent workflows under simulated crashes, network retries, and timeouts.
- Empirical Results: Tested across 1,000 deterministic execution seeds; hardened runtime achieved 0 invariant violations vs. 747 violations in naive unhardened agent loops.
- Stack: Python, Abstract Syntax Trees (AST), Deterministic Event Simulator, React Trace Viewer.
Codebase AST Visualizer & Native Claude Code Blast-Radius Skill
- Static code intelligence engine and native Claude Code Agent Skill mapping multi-language codebases into deterministic dependency graphs with zero cloud daemons or database setup.
- Blast-Radius Recall: Traces multi-hop upstream cascading impact in < 50ms via in-memory reverse-BFS, flagging CRITICAL, WARNING, and SAFE risk tiers.
- Archify-Grade Visualizer: Generates standalone single-file interactive SVG visualizers with glowing blast-radius paths and 1-click Mermaid export.
- Stack: Python 3.10+, AST Parsing, Directed Graph Engine, Reverse BFS, Claude Code Skill.
- AI & Retrieval: Retrieval-Augmented Generation (RAG), Hybrid Search (BM25 + Dense RRF), Neuro-Symbolic Agent Architectures, Conformal Risk Routing, Abstract Syntax Trees (AST).
- Backend & Systems: Python 3.12+, FastAPI, Pydantic v2, PostgreSQL, NumPy, REST APIs, Asynchronous I/O.
- Domain Standards & Interoperability: SEC EDGAR (HTML/XBRL), HL7 FHIR R4, JSON Schema, Docker, Git.