Past projects
and what they delivered.

Problem

Standard Windchill workflow templates could not support conditional branching based on part type, cost impact threshold, or affected assembly level. Approval routing was done manually by email outside the system, with no audit trail.

Solution

Built a custom WorkflowProcessor in Java that reads part classification attributes and change impact metadata to auto-route approvals. Implemented OperationDelegate validators blocking promotion without mandatory sign-offs. Added a server-side event listener posting ERP notifications on lifecycle state change.

Technology

Windchill 12.1 Java API, WTMethodServer, WorkflowProcessor, OperationDelegate, xconf configuration, Windchill event framework, REST webhook to ERP

Problem

No server-side validation enforced completeness before BOM release. Engineers could promote parts and BOMs with missing weight, material, or drawing references. Over 14,000 parts had at least one missing mandatory attribute.

Solution

Developed a set of server-side Windchill validators using ValueRequired constraints and custom OperationDelegate classes that block BOM promotion when mandatory attributes are absent. Built headless Java data utility scripts for bulk attribute repair, part deduplication, and BOM structure normalization.

Technology

Windchill Java API, OperationDelegate, ValueRequired constraints, WTPart/WTPartMaster subtypes, IBA attribute framework, headless Java batch utilities, Oracle JDBC

Problem

Windchill full-text search returned only metadata matches. The actual content of 600K+ PDF and Office documents was inaccessible. Engineers queried colleagues instead of the system, creating knowledge bottlenecks around senior staff.

Solution

Built a RAG pipeline that indexes Windchill document content into a vector database via the Windchill REST API and Apache Tika for document parsing. Engineers query in natural language and receive answers with citations linking back to source documents in Windchill.

Technology

Windchill REST API, Apache Tika, OpenAI GPT-4o, text-embedding-3-large, Pinecone vector database, FastAPI backend, React frontend, SAML SSO integration

Problem

No automation on the WhatsApp channel meant every message required manual reading and response. Order confirmations, delivery status requests, and standard product queries all consumed the same human attention as complex issues.

Solution

Built an AI agent on the WhatsApp Business API that classifies incoming messages, handles routine confirmations and status queries automatically, and escalates non-standard requests to the right person with full context.

Technology

WhatsApp Business API (Cloud), OpenAI GPT-4o, LangChain agent framework, webhook integration with ERP, message classification pipeline, Python FastAPI, Redis for session state

Problem

VOIP calls happened outside the CRM with no integration. Salespeople manually typed call summaries with low compliance. Pricing discussed, technical requirements, and delivery commitments never made it into the CRM.

Solution

Integrated VOIP webhook hooks to detect call start/end events and trigger automatic recording transcription. An AI pipeline processes each transcript to extract the entities (contacts, products, prices, commitments, next steps) and writes structured data back to the CRM via API.

Technology

VOIP webhook API (Twilio / 3CX), OpenAI Whisper for transcription, GPT-4o for entity extraction, CRM REST API (HubSpot / Salesforce), Python processing pipeline, PostgreSQL for call archive

Problem

No separation between dev and prod environments. API keys stored in plain text. A single Python process serving all requests with no queuing or load balancing. Model inference running on CPU with 8-second response times.

Solution

Designed and deployed a production infrastructure stack: Dockerized FastAPI services on Kubernetes, GPU node pool for inference, NGINX ingress with rate limiting, Redis queuing, Prometheus/Grafana monitoring, and a CI/CD pipeline via GitHub Actions.

Technology

Docker, Kubernetes (GKE), NGINX ingress, Redis, Celery workers, PostgreSQL, Prometheus, Grafana, GitHub Actions CI/CD, HashiCorp Vault, NVIDIA A100 GPU node pool, Terraform

Problem

Naive implementation passed the full conversation history to GPT-4o on every turn, causing exponential token growth, high latency, and monthly API costs that made the product economically unviable. TTS responses had 3–4 second delays.

Solution

Built a hybrid voice pipeline: OpenAI Whisper for STT, GPT-4o for reasoning with a custom memory layer, and Google Gemini 2.0 Flash TTS for low-latency speech synthesis. Implemented a sliding window memory system with semantic summarization of older turns.

Technology

OpenAI Whisper (STT), GPT-4o (reasoning), GPT-4o-mini (summarization), Google Gemini 2.0 Flash (TTS), LangChain memory with custom summarizer, WebSocket streaming, Redis session store