Portfolio

Three AI bots qualify leads in real time using Q0-Q4 framework. API health dashboard, token usage analytics, and OpenAPI docs. Streamlit BI dashboard with market pulse, commission tracking, and bot performance. GoHighLevel CRM integration. Agent cost tracker, RAG decision tracer, LLM observability, and conversation analytics.

Lead, Seller, and Buyer bots with structured Q0-Q4 question flows. Semantic analysis, temperature scoring, and cross-bot handoff with 0.7 confidence threshold.

Async-only LLM orchestrator with retry/backoff, per-request cost tracking, and structured JSON output. Claude + Gemini + OpenAI + Perplexity through one interface. ReAct agent loop, evaluation framework, model registry, agent memory, and multi-agent mesh. Tool registry with provider-format conversion, prompt template engine, token-bucket rate limiting. Pip-installable with CLI.

6 production-ready Model Context Protocol servers: SQLite CRUD, file organizer, TF-IDF knowledge base, system monitor, git analytics, task tracker. Analytics and cost tracking server. Auto-discovery CLI and Streamlit playground. 32 tools total.

Upwork job scanner with 105-point scoring rubric, 4-agent proposal pipeline, prompt injection testing suite (60+ patterns), LinkedIn outreach engine.

Upload CSV/Excel for auto-profiling, interactive dashboards, marketing attribution (4 models), anomaly detection (Z-score/IQR), predictive modeling with SHAP, time-series forecasting, K-means/DBSCAN clustering, feature engineering lab, model observatory (5 models), hyperparameter tuning, statistical testing, KPI framework, dimensionality reduction, advanced anomaly detection, and regression diagnostics.

25-page interactive finance dashboard answering 17 business questions: credit risk modeling (SHAP/RF), stress testing (CCAR/DFAST), VaR analysis, A/B testing, fair lending geography (real CFPB HMDA data), causal inference, and more. Delivered as dashboard, 8-tab Excel workbook, 11 SQL query templates, 15 Jupyter notebooks, and written business memos. dbt analytics layer (DuckDB + Snowflake).

Customer retention and delivery performance analysis on 100K+ real orders from a Brazilian e-commerce marketplace. Key findings: 93% of customers never return (vs 25-30% industry benchmark), late deliveries cost 1.2 review points, K-Means segmentation surfaces 4 actionable customer segments. Delivered as 10 SQL queries, EDA + segmentation notebooks, interactive dashboard, and written business memo.

Upload PDFs/DOCX, ask questions, get cited answers. Hybrid retrieval (BM25 + vectors), cross-encoder re-ranker, query expansion, retrieval evaluation (MRR, NDCG@K), 5 chunking strategies with comparison, citation quality scoring, redundancy detection, RAGAS-style evaluation, conversation manager, document graph, batch processing, JSON/CSV export.

Enterprise-grade retrieval-augmented generation with hybrid search (BM25 + dense vectors), reciprocal rank fusion, ChromaDB vector store, FastAPI REST API with health/metrics endpoints, and Grafana observability. Designed for principal AI engineer portfolio demonstration.

Production-grade DAG orchestrator with topological sort, retry/backoff, graceful degradation, BaseAgent ABC with Pydantic validation, ASCII monitoring dashboard.

YAML-configurable web scrapers, competitor SEO analysis with gap detection, asyncio scheduler with change callbacks, backlink estimation, data validation rules engine, diff visualization, keyword extraction (TF-IDF n-grams), content scoring, data quality profiler, content intelligence, and data pipeline.

8 built-in prompt patterns (CoT, few-shot, RAG, decomposition), TF-IDF evaluation (faithfulness, relevance, completeness), z-test A/B comparison with statistical significance, reproducible benchmark runner, 7 business categories with 35 sample tasks. Prompt evaluator with ROUGE and semantic similarity, prompt versioning, and safety checker.

Copy-pasteable LLM recipes: completion with FastAPI endpoint factory, SSE streaming, function calling with 3 demo tools, TF-IDF RAG pipeline, circuit breaker + retry hardening, per-request cost tracking, P50/P95/P99 latency measurement. Guardrails engine with injection detection, PII redaction, observability, and batch processor.

Full-stack AI dashboard with real-time metrics, chat interface, and document processing UI. Dark-mode design with Recharts analytics, shadcn/ui components, and responsive layout.

YAML-driven GHL bot framework supporting real estate, home services, and legal verticals. Configurable question flows, multi-vertical prompt templates, and GoHighLevel CRM integration.

YAML-defined async workflow engine with SSE progress streaming and 3 built-in example workflows. Background task queue with ARQ, real-time status updates, and webhook notifications.

Real-time voice AI pipeline: OGG to FFmpeg to VAD to STT to Claude to TTS, with sub-3s round-trip latency. WebSocket streaming, Silero-ONNX voice activity detection, Deepgram speech-to-text.

Drop-in AI chat widget with a single script tag. Full SaaS product: FastAPI + pgvector + RAG + WebSocket + Supabase JWT auth + Stripe billing. Shadow DOM isolation, streaming responses, Next.js admin dashboard. Live demo with no account required.

Production-ready Next.js 15 SaaS boilerplate with Supabase auth, Stripe billing, TypeScript, and Tailwind. Lazy Stripe initialization for edge-compatible builds.

Document extraction API with async job processing, pgvector semantic search, and ARQ workers. Extracts structured data from PDFs and text documents at scale.

Knowledge graph-enhanced RAG system. Entity extraction (spaCy + Claude), NetworkX graph traversal, BM25 + graph fusion retrieval, fact-checking, side-by-side comparison: Basic vs GraphRAG.