FREELANCE AI ENGINEER & DOCUMENT AUTOMATION.
I build pragmatic AI applications that solve real workflow bottlenecks without hallucinations or runaway cloud API costs. Specializing in high-volume OCR ingestion pipelines, vector search (Milvus), AST-grounded codebase RAG, and private local LLMs.
AI Engineering Solutions
Enterprise Document OCR & Ingestion
Asynchronous PDF, scan, and contract processing pipelines. Using RabbitMQ worker queues and Tesseract OCR to convert messy multi-page documents into structured, searchable text blocks.
- · High-concurrency worker pool parsing PDFs at scale
- · Semantic chunking preserving document headers & tables
- · Bounded memory usage preventing server out-of-memory crashes
Custom RAG & Vector Search
High-dimensional embedding pipelines with Milvus, Qdrant, or pgvector. Engineered with hybrid keyword-vector retrieval (BM25 + Dense) to return exact domain context instead of superficial fuzzy matches.
- · Re-ranking algorithms ensuring top-k precision
- · Role-based access control (RBAC) filtering at vector search time
- · Audit logging tracking all query context and model completions
Codebase AST Grounding (Zero Hallucination)
Eliminate model hallucinations when querying complex code repositories. By analyzing Abstract Syntax Trees (ASTs), model responses are strictly bound to verified file paths and line number evidence.
- · AST symbol extraction (functions, interfaces, structs)
- · Strict grounding verification rejecting unreferenced claims
- · Automated repository intelligence for technical recruiting and audits
Local-First Private AI (Ollama)
Deploy quantized open-weight models (Llama 3, Qwen, DeepSeek) on private on-premise hardware or edge devices. Keeps proprietary enterprise data 100% offline with zero cloud API billing.
- · Ollama daemon integration with Go/Python system background services
- · Zero internet network egress for strictly confidential operations
- · Sub-50ms token generation on modern GPU/NPU hardware
MetroMind & Klarity: Enterprise Intelligence in Action
Explore MetroMind (enterprise asynchronous OCR and vector search pipeline) and Klarity (evidence-grounded repository intelligence engine), demonstrating practical AI system engineering without vaporware.
AI Engineering FAQ
How do you prevent hallucinations in business RAG systems?
Most RAG failures occur from poor document chunking and naive similarity search. I implement strict syntactic grounding: the model is instructed to answer solely using citations with character or line-number evidence from retrieved chunks. If the retrieved evidence does not contain the answer, the system deterministically outputs an explicit non-match rather than hallucinating.
Can we run the models completely within our private VPC?
Yes. For clients dealing with sensitive financial, medical, or legal documents, I deploy local model engines using Ollama or vLLM directly on your own private cloud compute instances (AWS EC2, GCP Compute) with zero third-party API exposure.
Build Reliable AI Automation Today
Integrate document intelligence and vector search with proven, predictable engineering.
START AI PROJECT INTAKE