I build and maintain production GenAI applications — grounded conversational systems, agentic workflows, and observable pipelines.
Software engineer specializing in Python-based generative AI systems, conversational interfaces for enterprise operational data, and deterministic retrieval workflows. Hands-on experience developing LangGraph multi-agent DAGs, GraphRAG in Neo4j, MCP-style tool integrations, and full-lifecycle LLMOps with golden evaluation suites.
How My Background Maps to Disney DEEP&T
Proven experience building conversational AI platforms, grounding responses in enterprise source data, and maintaining resilient production GenAI services.
(:Stage)-[:PRECEDES]->(:Stage). Enforced a fail-closed grounding gate (is_grounded_in_neo4j returns null if 0 nodes match or confidence < 0.75). Verified Bolt Cypher query latency at ~90.07ms median local with an end-to-end <800ms target SLA.Production GenAI Systems
50 golden market vectors validate Pydantic coercion and null-math suppression (<800ms SLA).
9 investment styles evaluate fundamental risk and debt walls without calculating or hallucinating prices.
3PM Daily: Solo-Operated Market Closing Terminal
Between 1:00 PM and 3:00 PM PT, active market participants face auction prints (MOC) and earnings releases. Monolithic dashboards cause cognitive overload across hundreds of data points.
Engineered a 4-artifact pipeline: Ingests ~800 15-min bars into MySQL; runs TA-Lib C-bindings scanning ~827 symbols in 4.2s for 1x ATR stops; executes a LangGraph 9-agent DAG stress-testing fundamental moats; publishes 4 setups meeting Confluence Score ≥ +4 via Cloud Run + Redis.
Runs for ~$18/month in cloud infrastructure with >90% margin. 101 assertions across 50 golden vectors passing in 170ms in CI/CD. Zero LLM price hallucination permitted in client pipeline.
Lowenfeld stages (1947) modeled as property graph triples: (:Stage)-[:PRECEDES]->(:Stage). Graph structurally enforces chronological sequence.
Multi-node conditional DAG with explicit grounding gate (is_grounded_in_neo4j returns null below 0.75 confidence to prevent hallucinated badges).
Bolt Cypher ~90.07ms median local (50 runs). Mutation test caught 12 out-of-order failures upon edge deletion. Target SLA <800ms.
Artstache: Source-Backed Pedagogical GraphRAG System
Parents reviewing children's artwork at 6:15 PM face paper clutter and fatigue. Vector search conflates a 4yo tadpole with a 10yo comic, risking out-of-sequence pedagogical recommendations.
Modeled Viktor Lowenfeld's (1947) stages into Neo4j property graph triples. Built a 5-node LangGraph orchestration DAG in Python with an explicit grounding validation gate (is_grounded_in_neo4j) returning null below 0.75 confidence. Stripped all image EXIF data in an <80KB in-browser Canvas buffer designed with COPPA privacy principles in mind.
Bolt Cypher ~90.07ms median local (50 runs, raw: latency_runs.jsonl). [:PRECEDES] graph relationships structurally prevent out-of-order milestones (50/50 passed; mutation test caught 12 failures). No-Line Rule paper elevation system.
Clean editorial landing for 30s scanning (0.00 CLS median, 15.56ms cascade); deep interactive tool sandbox isolated in /lab.
40/50 queries trapped client-side in browser memory with $0 LLM cost and <10ms P50 latency across 2,500 test runs.
Python DCF model runs in 3.63ms median with 92% token savings (1850→140). Local watchdog daemon warms database pools.
don-the-imaginator.com: From Gated Friction to Open IA
Mandatory Firebase auth gate created severe drop-off: Firestore SiteAnalytics revealed Funnel: 20% signup, 0% activation (N~10/mo). Exclusivity hypothesis failed; technical evaluators demand immediate proof without friction.
Dismantled the auth gate. Bifurcated information architecture into a reserved editorial landing for rapid review and an isolated `/lab` sandbox for technical diligence. Built client-side Pocket-Synth trapping 80% of queries locally ($0 LLM cost), paired with a FastAPI backend executing deterministic Python tools (DCF in 3.63ms median) and generative React cards.
Zero mandatory auth friction with immediate <5s proof. Verified Test 1 (Theme Engine, 250 swaps): 0.00 CLS median, 15.56ms cascade. Local watchdog daemon maintains warm pools (disclosed limitation; next: Cloud Scheduler).
How I Build Resilient GenAI Systems
Rigorous production standards from schema design and deterministic grounding to continuous observability and regression testing.