Portfolio 

Two independent XGBOOST bundles power the predictive layer: an APPLICANT LATE-PAYMENT MODEL (15 features, 16k samples — AUC 0.784, Brier 0.135) and a 24-HEAD RESIDENT RISK MODEL spanning six families (late / frequency / severity / arrears / cure / retention). NO LLM sits anywhere in the scoring path, so a score is a pure function of its features and reproduces exactly in a test. The agents only narrate the numbers — every prediction ships with SHAP-STYLE REASON CODES and a confidence interval.

Two retrieval systems answer two different question shapes. LLAMAINDEX RAG handles lease documents — hybrid dense + lexical search with FlashRank reranking over CHROMADB or PINECONE, answered with citations back to the source clause. Graph-RAG handles relationship questions by compiling natural language to TEXT-TO-CYPHER against a real NEO4J property graph — executed read-only under RoutingControl.READ, enforced by the driver itself rather than a keyword blocklist, so a crafted question cannot mutate the graph.

The evaluation harness is the centerpiece, and it GATES CI: 300 golden items are graded twice — once by a deterministic layer (eligibility accuracy 100%, extraction field accuracy 100%, recommendation NDCG@5 85%, plus routing, grounding and safety all at 100%) and once by a RAGAS tier (faithfulness 100%, groundedness 93%, eligibility-explanation consistency 100%) — and GitHub Actions fails the build on regression. A head-to-head prompt A/B even picked its winner on measured groundedness 100% vs 85% while being both faster and cheaper. The scoring tier itself is ADVERSARIALLY VALIDATED before it is trusted, so every grounded answer is verified against source facts rather than assumed.

Tech: FastAPI • Pydantic v2 • React 18 • TypeScript • Vite • LangGraph • LlamaIndex • neo4j-Graph-RAG • Claude (Sonnet) • XGBoost • scikit-learn • ChromaDB / Pinecone • Neo4j • Neon Postgres • RAGAS • LangSmith • Playwright • Vitest • Pulumi

This is not just RAG. It is a controlled, production-shaped system for answering questions over real SEC 10-K / 10-Q filings with VERIFIABLE CITATIONS.

The workflow is intentionally ULTRA OPTIMIZED for latency + cost: a DETERMINISTIC ROUTER classifies intent (single-company, comparison, ingestion) with NO LLM CALL, then a FAST RAG AGENT compresses retrieval + analysis + reporting into ONE LLM call (about 3× fewer calls than traditional chains). The payoff is measured: repeated queries return 485× faster (9.69s → 20ms), and a two-stock comparison lands in 16ms against a 12.21s baseline.

Beyond vector search, a Graph-RAG SERVICE backed by NEO4J AURA (with a zero-dependency in-memory store as a drop-in alternative) injects a RELATIONSHIP BRIEF as a first-class retrieved chunk — so peer, supplier, and cross-company questions get reasoned relationship context, not just nearest-neighbor text. Numeric questions can additionally pull authoritative figures from the SEC XBRL COMPANYFACTS API.

Retrieval runs on PINECONE (SERVERLESS) with HYBRID RETRIEVAL, SECTION BOOSTING, CROSS-ENCODER RERANKING on by default, and MULTI-QUERY EXPANSION that broadens financial vocabulary ("operating expenses" → R&D / S&M / G&A). A boilerplate filter strips tables of contents, auditor reports, and signature pages before ranking, a widened candidate pool (top_k × 6) feeds the reranker, and a recency layer prefers the newest 10-K for annual questions while still reserving slots for highly-relevant older disclosures. An MD5-KEYED SEMANTIC CACHE returns repeated queries in ~20ms.

Ships with a MULTI-PROVIDER LLM abstraction (Anthropic Claude Sonnet 5.0 / OpenAI / local Ollama) with pooled + cached instances, plus performance engineering via async FastAPI, NDJSON response streaming, request deduplication, and concurrent comparison execution. Chat history persists to NEON POSTGRES on a best-effort path that never blocks an answer.

Quality is not asserted, it is MEASURED. An LLM-AS-JUDGE harness scores a 250-QUESTION GOLDEN SET on faithfulness, relevance, and cross-company attribution: 100% factual accuracy (49/49), 4.81 / 5 faithfulness, ZERO hallucinations, ZERO missing citations, zero ticker-isolation violations, and 12/12 adversarial prompts correctly declined. Prompt engineering on grounding alone lifted faithfulness from 4.30 to 4.83. Testing spans pytest (unit + integration), Vitest (frontend), and Playwright (end-to-end).

A production-style policy compliance assistant that answers questions from internal policy documents with REAL CITATIONS, REAL-TIME STREAMING, and optional MULTIMODAL (text + images) support using CLIP library.

The RAG pipeline is orchestrated with LANGGRAPH + LANGCHAIN and follows a clear flow (query embedding → vector search → top-k retrieval → context assembly → LLM generation → citation extraction). Retrieval runs on PINECONE (SERVERLESS) with semantic search + metadata filtering, enhanced with QUERY EXPANSION, HYBRID SEARCH (semantic + keyword), and relevance boosting via RERANKING using a CROSS-ENCODER(plus optional AUTO REWRITE to refine user queries). A REDIS caching layer reduces repeated retrieval/LLM calls for faster, more consistent answers. Supports a unified MULTI-PROVIDER interface (Ollama, OpenAI, Anthropic) and streams token-by-token with SSE. Every answer includes citations + is recorded to POSTGRESQL AUDIT LOGS for traceability; shipped with a production-friendly DOCKER COMPOSE setup.

An AI travel planner that turns a city and a trip length into a day-by-day itinerary with real street addresses, per-day route maps, and live Google ratings for every stop. Generation is FANNED OUT ACROSS CONCURRENT CLAUDE SONNET 5.0 CALLS — one for the trip overview, then one per group of days — and STREAMED TO THE BROWSER OVER SSE as each chunk lands.

Plans are enriched with live Google Places data — star ratings, review counts, photos, and coordinates resolved per stop and cached for the session, so switching day tabs issues zero additional API calls.

Tech: Next.js 14 (App Router) • React 18 • Tailwind CSS • DaisyUI • TanStack Query • Prisma + PostgreSQL • FastAPI + Uvicorn • Python 3.11 • Anthropic Claude Sonnet 5.0 • Google Maps & Places • Unsplash • Redis + cachetools • orjson • Docker Compose • Nginx • AWS EC2 + CloudFront • Pulumi (TypeScript)

Enterprise Scale project intentionally built to showcase AGENTIC AI patterns inside a real e-commerce app. The React frontend talks to an AGENT GATEWAY through a single chat surface. The agent performs INTENT ROUTING and delegates to either an internal MCP TOOL SERVER (explicit tools for product search/lookup + order tracking) or a RAG SERVICE (local docs retrieval for policy/support Q&A).

This is a production-style full-stack e-commerce application built using React, FASTAPI, MONGODB, and DOCKER. It provides a complete end-to-end shopping experience—from user authentication and product browsing to secure checkout and order tracking—all backed by scalable async REST APIs and ROLE-BASED admin management.

The system includes a fully functional admin dashboard for managing users, products, orders, and inventory, along with secure JWT-based authentication, image upload support, and PAYPAL + STRIPE payment integration. 🛡️ The frontend communicates with a containerized backend API and MongoDB database, forming a deployable, cloud-ready full-stack solution.

The system understands intent and automatically chooses the correct backend action. 🤖 The key production-shaped boundary is that the UI never calls TOOLS or RETRIEVAL directly—everything stays behind the GATEWAY—making the tool-use layer AUDITABLE and TESTABLE, while enabling optional HYBRID SEARCH (BM25 + embeddings) via PINECONE/OPENAI.

A reproducible, end-to-end fine-tuning pipeline for building a domain-specialized POLICY COMPLIANCE model using QLORA (4-bit quantization), then deploying it for local inference with OLLAMA. Built for repeatable experiments and real deployment, from dataset → fine-tune → packaging.

Implements parameter-efficient fine-tuning of LLAMA 3.1 8B with 4-BIT NF4 quantization—enabling training on a SINGLE GPU. Policy answer accuracy improved by about POLICY ANSWER ACCURACY IMPROVED BY ABOUT ~70% versus the base model, with strong convergence (training loss reduced by ~79% over 3 EPOCHS). Includes an automated EVALUATION HARNESS, a clean END-TO-END PIPELINE (data generation → fine-tuning → adapter merge → evaluation → packaging), and config-driven experiments with clean artifact separation. Ships a merged model packaged for OLLAMA (adapter merge → GGUF) to support fast, OFFLINE INFERENCE.

Tech: Python • HuggingFace Transformers • PEFT (LoRA / QLoRA) • BitsAndBytes • Accelerate • Ollama • GGUF

A production-ready FINE-TUNING and evaluation pipeline for GPT-4O-MINI, designed to improve enterprise policy compliance accuracy while reducing TOKEN USAGE and LATENCY. The domain-specialized model trained on 164 curated JSONL examples boosted accuracy from 35% → 77.5% (~2.2× better) on compliance-focused evaluation sets. Includes an AUTOMATED EVALUATION SUITE with CATEGORY-LEVEL SCORING, REGRESSION CHECKS, and performance reports—token usage dropped by ~20%, and response latency improved from 1.61s → 1.06s (~44% faster). The pipeline follows an MLOPS-STYLE WORKFLOW with training data versioning, job automation, and model validation in OPENAI PLAYGROUND.

A fully SERVERLESS, EVENT-DRIVEN Order Management System on AWS, built for high scalability, operational visibility, and real-time analytics. Order workflows run through AWS LAMBDA behind API GATEWAY, with decoupled processing via SQS, SNS, and DYNAMODB.

Complex lifecycle flows are coordinated with STEP FUNCTIONS (retries, failure paths, DLQs) and optional EVENTBRIDGE pipelines. The whole environment is provisioned with AWS CDK (TypeScript) + CloudFormation, and shipped with a modern REACT + TYPESCRIPT dashboard featuring analytics charts (Chart.js) and CSV import/export workflows. Security and observability lean on least-privilege IAM + CloudWatch monitoring.

Tech: AWS Lambda • API Gateway • SQS • SNS • DynamoDB • S3 • Step Functions • EventBridge • AWS CDK • React 19 • TypeScript • Chart.js

Mixmaster is a full-stack cocktail discovery app built with React ⚛️, React Query ⚡, Node.js 🌐, and PostgreSQL 🗄️. It blends real-time data from TheCocktailDB 🍸 with custom user-created recipes stored locally with uploading your image to S3 bucket🪣, enhanced by intelligent caching for high-speed performance. A serverless AWS Lambda microservice ☁️ handles newsletter automation with Amazon SES ✉️, delivering a smooth, modern UI and fast, reliable CRUD operations.

A fully Dockerized MERN e-commerce platform built with React ⚛️ + Redux Toolkit 🧩, Node/Express 🚀, and MongoDB 🗄️. It delivers a real-time product experience with secure JWT authentication 🔐, PayPal/credit-card checkout 💳, image uploads 📸, and full admin management workflows. Designed with a scalable, production-ready architecture ideal for real-world store operations.