Key Skills

Technical Skills

• Languages & Frameworks: Python 3.12, SQL (Multi-Dialect OLAP & OLTP), TypeScript 5.x, Next.js 16 (App Router), React 19, FastAPI, LangGraph, LangChain, Pydantic v2, sqlglot, HTML5/CSS3, Tailwind CSS
• Cloud Data Warehouses & OLAP Engines: Snowflake (Multi-cluster shared data, Time Travel, Resource Monitors), Google BigQuery (Dremel serverless, BI Engine, partitioned/clustered ELT), AWS Redshift Serverless (RPU capacity management, distribution/sort keys), Databricks SQL (Photon vectorized lakehouse, Delta Lake, Unity Catalog), ClickHouse (MergeTree / ReplacingMergeTree, SIMD real-time analytics), DuckDB (In-process vectorized OLAP), Apache Iceberg
• Cloud Computing & Platforms: Google Cloud Platform (GCP — Cloud Run, Vertex AI, Cloud Pub/Sub, BigQuery Log Analytics, Cloud Firestore, Cloud Logging & Monitoring, IAM, Cloud Storage), Cloudflare Edge (TLS 1.3, HTTP/3, DNS CNAME Flattening), AWS (Redshift Serverless, S3)
• Infrastructure as Code (IaC) & DevOps: Terraform, Docker, Docker Compose, Git, GitHub Actions, Multi-Dialect SQL CI (sqlglot + Pytest + Ruff), Pre-commit/Pre-push Hooks
• Agentic AI & LLMs: Google Gemini 1.5 Flash & Pro, Vertex AI Gemini (Multi-turn tool calling), LangGraph (Stateful multi-actor workflows), LangChain, Truth-Preserving Prompt Engineering, Google XYZ Metric Synthesis, In-Memory BM25 RAG, Ollama
• Databases & Vector Stores: PostgreSQL 16 (pgvector 768d dense embeddings, Cosine Distance indexing), Cloud Firestore, BigQuery, Snowflake, ClickHouse, DuckDB, SQLite, MySQL, MongoDB, Redis
• System Performance & Concurrency: Async Threadpool Decoupling (loop.run_in_executor), Cold vs. Warm OLAP Query Benchmarking, Route Prefetching (<45ms), React.memo Profile Optimization, Sub-second Heuristic Fallback Engines
• Security & Cryptography: Native Node.js crypto.scrypt (Memory-hard key derivation), Constant-time comparison (crypto.timingSafeEqual), Prompt Injection Output Guardrails (checkReply), Rate Limiting, Cryptographic Commit/Actor Verification
• Site Reliability & Observability: Cloud Monitoring, Cloud Logging, Incident Triage & RCA, Chaos Simulation, FinOps Scale-to-Zero & Warehouse Spend Guardrails
• Machine Learning & Statistics: Linear/Logistic Regression, Ensemble Trees, Gradient Boosted Trees, ANN, CNN, LSTM

Cloud Data Warehouse Comparison & Multi-Engine OLAP Benchmarking

I architected and built the Cloud Data Warehouse Comparison & Benchmarking Suite, evaluating Snowflake, Google BigQuery, AWS Redshift Serverless, Databricks SQL, ClickHouse, and DuckDB across standardized 2M-row analytical workloads:

• 5-Archetype OLAP Benchmark Suite (Q1–Q5): Benchmarks high-selectivity scans, multi-column GROUP BY aggregations, time-series rollups, star-schema joins, and partitioned ROW_NUMBER() window functions over 2,000,000 NYC TLC Yellow Taxi records, tracking cold vs. warm latency, bytes scanned, and cost per 1,000 queries.
• Multi-Dialect SQL AST Verification (sqlglot): Automated CI pipeline parsing native SQL dialects (bigquery, snowflake, redshift, clickhouse) into Abstract Syntax Trees to verify query structure and enforce FinOps cost guardrails (AUTO_SUSPEND = 60, maximum_bytes_billed, Base RPU = 8).
• Data Governance & Table Format Architecture: Comparative evaluation of ACID isolation, constraint enforcement, Time Travel retention (up to 90 days in Snowflake), column-level lineage (Unity Catalog / Horizon), and open lakehouse formats (Apache Iceberg, Delta Lake).

View GitHub Repository →

HireCraft — Closed-Loop AI Platform & Vector Search Architecture

I architected and built HireCraft, a technical job discovery and AI resume tailoring platform integrating Next.js 16, React 19, FastAPI, and PostgreSQL with pgvector:

• 768d Dense Vector Semantic Matching: Projected candidate profiles and verified job descriptions into 768-dimensional vector spaces using pgvector, achieving <18ms Cosine Distance queries directly in PostgreSQL and reducing vector infrastructure costs by 90% compared to Pinecone/Milvus.
• Truth-Preserving Google XYZ Synthesis: Designed strict architectural constraints eliminating LLM hallucinations, refactoring candidate achievements into verified Google XYZ format (Accomplished [X], measured by [Y], by doing [Z]) with interactive fact prompting for unquantified bullets.
• Asynchronous AI Gateway Concurrency: Resolved a 10-second FastAPI event-loop freeze by offloading synchronous Google Gemini API calls into a threadpool executor with 4.0s timeouts and sub-second (<20ms) heuristic fallbacks.
• Sub-50ms Navigation & ATS Compliance: Reduced client navigation latency from 1.8s to <45ms (97% drop) via Next.js route prefetching and React.memo capacity budgeting, with clean @media print styling delivering a verified 96.4% ATS machine parse rate.
• Native Cryptographic Security: Hardened user auth using Node.js native crypto.scrypt memory-hard key derivation and crypto.timingSafeEqual, eliminating third-party C++ binary dependencies (bcrypt).

Cloud SRE Copilot & Autonomous Multi-Agent Operations (GCP)

I architected and built the Autonomous Cloud SRE Copilot & Incident Investigator, a production-grade multi-agent operations showcase on Google Cloud Platform:

• Agentic Multi-Turn Tool Calling: Integrated Vertex AI Gemini 1.5 with custom GCP diagnostic tools (query_cloud_logging, get_metric_timeseries, query_log_analytics, inspect_gcp_resource) enabling the agent to autonomously formulate hypotheses, query live telemetry, and isolate root causes.
• Event-Driven Incident Pipeline: Engineered an end-to-end telemetry pipeline routing Cloud Run exception logs through Cloud Logging Sinks and Cloud Pub/Sub with Dead-Letter Queues (DLQ) to trigger real-time webhook triage.
• Chaos Engineering & Auto-Remediation: Built a failure simulator covering Cloud Run misconfigurations, IAM drops, OOM container kills, and connection starvation, with automated zero-downtime rollback remediations guarded by human-in-the-loop authorization.
• Modular Terraform & FinOps: 100% codified in reusable Terraform modules (iam, messaging, observability, cloud_run, storage_data) with least-privilege service accounts and scale-to-zero serverless architecture achieving $0/month idle standby cost.

View GitHub Repository →

LangGraph Multi-Actor Agent Engineering

I architected and built the LangGraph Playground Monorepo, a modular engineering workspace demonstrating production-ready stateful agent patterns:

• Stateful Cyclic Workflows & Reducers: Designed core state schemas (BaseAgentState, SQLAgentState) utilizing message reducers (add_messages) for continuous context passing.
• Conditional Routing & Tool-Calling: Built dynamic agents with multi-branch routing between model inference and custom tools (calculator, SQL utilities).
• Session Checkpointing & Memory: Integrated MemorySaver checkpointers for persistent multi-turn conversational agents.
• Self-Healing SQL Generation: Engineered guardrailed agents that autonomously catch SQL execution errors, retry with corrected syntax, and block destructive operations (FORBIDDEN_SQL_KEYWORDS).

View GitHub Repository →

ETL Pipeline Skill Suite

I designed a set of reusable, composable AI agent skills that encode an entire data team’s pipeline conventions — directory structure, coding standards, and validation rules — so scaffolding a new data pipeline or auditing an existing one follows a consistent, self-checking process instead of relying on tribal knowledge.

What it does: I route each request — build a new pipeline, fix an existing one, or extend one with a new data category — to the right specialized skill. Each skill scaffolds or audits the standard pipeline stages — data collection, cleaning/transformation, and output — including a genuinely detailed document-extraction path I built that handles both text-based and image/scanned document sources, with automatic page detection and OCR fallback. I added a companion skill that generates and audits data-validation schemas, and another that generates the pipeline’s run configuration (schedule and step dependencies). I gave every skill the same built-in safeguards: it won’t write changes on a protected branch without confirmation, it supports both a guided step-by-step mode and a fully automated mode, and it lets me navigate backward to revise an earlier decision without losing downstream work.

Why it matters: I shifted pipeline scaffolding from “an engineer manually follows a checklist” to “an agent enforces the checklist automatically,” while keeping a human in the loop for any decision that touches production data or a shared branch.