About Projects Core

Portfolio

Projects

Production builds — enterprise AI pipelines, agentic tools, full-stack platforms, and hardware ML. Each window below is a live demo or animated walkthrough.

01

Enterprise Agentic RAG

An AI assistant that answers questions from a company's own documents — and, like a careful researcher, checks and corrects its own answer before replying.

Under the hood

Production-grade agentic RAG reference architecture. A LangGraph pipeline (planner → retriever → synthesizer → critic) with self-correction loops, hybrid retrieval combining pgvector + BM25 + RRF, cross-encoder reranking, OpenTelemetry + Prometheus observability, per-request cost tracking, and full CI/CD with Docker Compose and GitHub Actions.

LangGraphAmazon Bedrockpgvector FastAPIDockerGitHub Actions
enterprise-agentic-rag — LangGraph Agentic Pipeline
Planner — query rewriting & decomposition
↓ structured sub-queries
Retriever — pgvector + BM25 + RRF re-rank
↓ ranked chunks + citations
Synthesizer — Claude / Bedrock generation
↓ grounded answer
Critic — self-correction + RAGAS eval
Recall@k · MRR · NDCG OpenTelemetry · Prometheus Cost per request tracking
02

Deep Research Agent

Give it any topic and it researches the web for you — reading multiple sources and writing a clear, cited report, like an automated research assistant.

Under the hood

Interactive CLI research agent built on the Strands Agents SDK and Amazon Bedrock. Runs a multi-tool pipeline — web search, page extraction, and site crawl via Tavily — to synthesize comprehensive, cited Markdown reports. Streaming output, partial-save on interruption, session management, and rotating logs.

PythonStrands Agents SDK Amazon BedrockTavilyClaude 3.7
deep-research-agent — Strands · Bedrock · Tavily
$ uv run python -m deep_research_agent
✓ Session f3a9c2 · Claude 3.7 Sonnet via Bedrock
query > "impact of AI on drug discovery 2025"
[tavily_search] "AI drug discovery 2025" → 10 results
[tavily_extract] nature.com/ai-pharma → 4,218 tokens
[tavily_crawl] deepmind.com/alphafold → depth 2
## AI in Drug Discovery — 2025 Overview
AlphaFold 3 cut target-ID time by 73% vs. prior baselines...
Novel candidates now proposed in <48 h end-to-end...
Sources: [nature.com] [deepmind.com] [nih.gov]
Output → research-output/ai-drug-discovery_20260525.md
03

SAP ABAP Flight Booking Platform

The complete behind-the-scenes engine for a flight-booking site — search flights, book a seat, pay, cancel, get a refund, and receive an e-ticket — built on the enterprise software big companies actually run on.

Under the hood

Full-stack flight booking system on SAP ABAP OO / HANA. Facade pattern over 8 domain classes (flight, booking, customer, payment, ticket, validation, cancellation) exposed through 4 REST HTTP handlers. CDS view data access layer over SAP standard flight tables with full booking lifecycle — search, reserve, pay, cancel, refund, and e-ticket generation.

SAP ABAP OOSAP HANACDS Views REST APIsFacade Pattern
sap-abap-workspace — ABAP OO · SAP HANA · REST
REST Handlers
GET /searchPOST /book PUT /manageGET /admin
↓ ZCL_FACADE_1986
Domain Classes
FlightBookingCustomer PaymentTicketValidationCancellation
↓ CDS Views
Data Access Layer
ZI_FLIGHT_1986ZI_BOOKING_1986 ZI_CARRIER_1986ZI_CUSTOMER_1986
↓
SAP HANA
SCARRSPFLISFLIGHT SBOOKSCUSTOM
04

Mobile Mechanic Platform

A website where customers can check if a mechanic serves their area and book one to come fix their car at home — try the live demo on the right.

Under the hood

Full-stack booking platform for an on-demand mobile mechanic service. Next.js frontend with service area lookup and appointment scheduling, Python backend with Supabase/PostgreSQL, Playwright end-to-end test suite with GitHub Actions CI, deployed on Vercel.

Next.jsTypeScriptSupabase PostgreSQLPlaywrightVercel
mobile-mechanic-website-cetb.vercel.app
05

AI-Powered SOW Editor

Paste in a meeting transcript and it auto-drafts the project contract, while an AI assistant suggests the rest — you just accept, reject, or edit, then export to Word or PDF.

Under the hood

React-based Statement of Work editor with real-time Claude AI assistance via AWS Bedrock. Rich text editing with Slate.js, inline AI suggestions, document state management with Zustand, change tracking, undo/redo, and PDF/DOCX export.

ReactSlate.jsClaude AI AWS BedrockZustand
sow-editor — Transcript → SOW → Export
① Meeting Transcript
"...the client is Acme Corp, migration by Q3,
budget around $200k, deliverables include..."
↓ entity extraction
② SOW Template
Client: Acme Corp
Timeline: Q3 2025
Budget: $200,000
Risk section: ___
↔
③ Claude Chat
"Add a data sovereignty risk clause for EU compliance?"
Accept Deny Edit
↓ finalize
④ Export DOCX PDF
06

Gesture Music Controller

Wave your hand in the air to control Spotify — skip a song, pause, or change the volume — using a tiny motion sensor and machine learning. Watch the video demo on the right.

Under the hood

Real-time mid-air gesture recognition for Spotify control. An ESP32 + IMU sensor streams 6-axis data over serial; Butterworth DSP filters clean the signal before a hybrid ML + physical-polarity classifier achieves 99%+ accuracy. A live 6-panel Matplotlib dashboard shows waveforms, confidence, and Spotify metadata in real time.

PythonC / ESP-IDFScikit-learn ESP32Matplotlib
youtube.com — Gesture Music Controller Demo
Gesture Music Controller Demo
▶
07

Real-Time Voicebot

A voice assistant you can talk to naturally — and even interrupt mid-sentence, just like a real phone conversation, with answers in under half a second.

Under the hood

Voice assistant powered by Amazon Nova Sonic via AWS Bedrock's bidirectional streaming API. Features barge-in detection (interrupt mid-response like a real conversation), async tool use for live data lookup, a Streamlit UI, and PyAudio-based I/O with separate 16 kHz input and 24 kHz output streams.

Amazon Nova SonicAWS Bedrock PyAudioStreamlitFastAPI
voicebot — Amazon Nova Sonic · Bidirectional Streaming
● LIVE — Nova Sonic Stream
Bidirectional Audio · Barge-In Enabled
Latency: <500msCost: <$0.03/session
08

AI News Video Generator

Automatically turns the day's trending news stories into short narrated videos — writing the script, generating the visuals, and rendering the final clip with no human editing.

Under the hood

Automated pipeline that scrapes trending articles from BBC and CNN, generates video scripts with GPT, and renders MP4s with AI-generated images and text overlays. Fully customizable video duration, pacing, and content sources — runs end-to-end without human intervention.

PythonOpenAI GPT Web ScrapingVideo Gen
news-video-generator — OpenAI GPT · BBC · CNN
AI News Video Generator — Live Feed

Fed Signals Rate Cut Amid Cooling Inflation Data

BBC · Generating script... ██████░░░░ 65%

Breakthrough in Protein Folding Opens Drug Discovery

CNN · ✓ Rendering video

Global EV Adoption Hits Record 18M Units in 2025

BBC · Queued