AI Engineering
I build practical AI systems — LLM applications, RAG, agents, document intelligence and automation — wired into business data, tools and workflows, and deployed like any other production system.
01Not a chatbot on a website
I build practical AI applications — systems where a model is one component inside a real product, wired to business data, tools and workflows — rather than adding a chatbot to a website and calling it AI.
A capable model with no retrieval, no tools and no validation is a demo. The application layer — context, guardrails, evaluation, cost control — is where reliability comes from.
Answers and actions should come from your documents, records and systems, with sources attached, not from whatever the model remembers from the internet.
When output feeds code, it is schema-constrained and validated. Free text is for humans; JSON is for systems.
Automation that can send, spend or delete gets an approval step. The goal is trust, not just autonomy.
02Architecture
From the model outward: each layer is engineering work, and each is where projects succeed or fail.
AI model
Choosing hosted or open models per task, with provider abstraction so the choice can change.
Application layer (the harness)
Prompt and context management, output validation, evaluation sets, logging and cost controls.
Business data
Ingestion pipelines, chunking, embeddings, vector and keyword indexes, access control on retrieval.
Tools
Well-typed functions the model can call — your APIs and internal services, with limits and audit logs.
Workflows
Queues, retries, approvals and human-in-the-loop steps that make automation trustworthy.
User interface
Interfaces that show sources, allow correction and fit how people already work.
Deployment
Containers, CI/CD, secrets management, monitoring of latency, quality and spend.
03Capabilities
Applications built around modern language models.
Prompt design, context management, evaluation and cost control for features that run on language models in production.
Knowledge bases, document search and semantic retrieval.
Chunking, embeddings, hybrid retrieval and reranking so answers are grounded in your own documents and data.
Tool calling, APIs, workflows and multi-step tasks.
Agents that call your systems through well-defined tools, with guardrails, retries and human checkpoints where they matter.
AI-assisted business processes and repetitive workflow automation.
Classification, routing, drafting and enrichment steps embedded inside existing queues, CRMs and back-office flows.
Extraction, classification, summarization and document question answering.
Turning PDFs, forms, emails and scans into structured data your systems can validate and act on.
Semantic search and intelligent knowledge retrieval.
Search that understands intent across products, records and documentation — combined with keyword search where precision matters.
Reliable JSON outputs, classification, extraction and workflow integration.
Schema-constrained outputs, validation and fallbacks so model responses can be trusted by downstream code.
Connecting models to business data, applications, APIs and internal systems.
Model providers, vector stores and internal APIs integrated into applications with observability and access control.
04Topics
The kinds of systems I design and build — each one a different engineering problem, not a different prompt.
Products and features where a language model does real work — drafting, answering, transforming — with the engineering around it to make that reliable.
Retrieval-augmented generation over your own documents and records: ingestion pipelines, chunking strategy, embeddings, hybrid search and citation.
Search that understands meaning rather than exact words, layered on top of keyword search so precision is not lost.
Multi-step systems that plan, call tools and complete tasks against your APIs, with limits, logging and human approval where risk is high.
AI steps embedded in existing business processes: triage, classification, extraction, enrichment and routing.
Extracting structured, validated data from PDFs, forms, invoices, emails and scanned documents.
Internal knowledge bases that staff and customers can actually query, with access control and source attribution.
Assistants grounded in your data and connected to your systems — able to look things up and take actions, not just chat.
Generation pipelines with templates, review steps and structured inputs so output stays on-brand and accurate.
Schema-constrained outputs (JSON, enums, typed fields) that downstream code can trust and validate.
Connecting hosted and open models into existing applications with provider abstraction, caching, cost and latency control.
End-to-end flows that combine models, rules, data and people — designed around how the business operates.
05Harness engineering
A harness is everything in an AI agent except the model — the tools it can call, the context it sees each turn, the memory that survives a session, the permissions and budgets that bound it, and the evaluations and traces that show whether it worked. A harness engineer designs and maintains that layer so an agent is reliable in production, not just in a demo.
Tool contracts with typed inputs and outputs, skills and instructions that encode how work is done here, context assembled per step rather than an ever-growing transcript, and approval tiers for risky actions.
Structured-output validation, second-pass verification, regression scripts against real data, tests and linters the agent must pass, and traces with cost per run.
Per-agent budgets with hard stops, permission scopes and sandboxes, human checkpoints for writes, and kill switches — shipped off by default and turned on deliberately.
When an agent fails, the fix goes into the harness — a validator, a tool, a rule, a test — so the same mistake cannot recur. Prompt tweaks are the last resort.
Built this way: an AI project manager inside a task platform with per-project memory, budget hard stops and an approval queue; Claude Code skills that package SEO and study-material workflows as reusable harness components; RAG assistants with PII masking, self-hosted reranking and evaluation passes.
What a harness engineer doesWriting
AG-UI is an open event protocol for streaming an AI agent to a frontend. What it is, how a run looks on the wire, how it differs from MCP, and when to skip it.
A VITE_ or NEXT_PUBLIC_ variable ends up in JavaScript every visitor can read. Here is how an AI key gets into a bundle, how to check your own app safely, the server-side fix, and what to do if it leaked.
A bot check decides who gets in, not how much they spend. A source-checked look at denial of wallet, what OpenAI and Anthropic do and do not cap, and a per-user budget design with a reserve and settle sketch.
From retrieval and agents to document intelligence, I build AI systems that plug into real business data and workflows.