About

Ami Heines — agentic AI systems

I build AI agents that do real work inside a business. They read your data, use your systems, and finish tasks that people now do by hand. I also build the controls that make them safe to run: permissions checked in code, a person approving anything that moves money or changes a record, tests on every change, and hard limits on spending.

The short version

I’m a senior software architect with thirty years in software, from developer to team leader running small teams. For most of that time I built the internal systems businesses run on: management software for construction companies, salary processing, CRM customization, an automated lending platform, and ticket sales for a lottery operation. From 2019 to 2026 I was backend architect and team lead for a security-compliance (GRC) platform.

That history is the reason I build agents. An agent is only useful when it is connected to the systems where the work happens, and most of the engineering is in that connection: the data model, the permissions, the business rules, and what happens when something fails. I have been building those systems for thirty years.

I deliver in one of two ways: by leading a small, senior remote team (at CISOteria, developers across Israel and India), or by building it myself with AI development tools like Claude Code, so one experienced architect ships at the pace of a small team.

Recent work

  • A LangGraph banking copilot (personal build, synthetic data): an intent router, retrieval, a tool-calling agent, and a verification step. Every action that moves money stops for human approval, and the decision is saved as an audit record. Before a reply reaches the customer, code checks every number in it against the tool results. Code and tests: github.com/smallestbusiness/banking-copilot.
  • A fine-tuned intent router for that copilot (personal build). It replaced an LLM call and was both more accurate — 94.2% against 85.3% for Claude Sonnet 4.6 on the BANKING77 benchmark — and much faster: 12 ms per message instead of 2 seconds. Code and results: github.com/smallestbusiness/intent-router.
  • Search over all of English Wikipedia: 35 million chunks in Postgres with pgvector, about 1 ms per search, with an agent mode that plans several searches before it answers. Public demo at ask.amiheines.com.
  • Guardrail, a gateway that stops AI apps and agents at a hard spend limit, so an agent stuck in a loop cannot run up the bill. Live at guardrail.smallestbusiness.com.
  • A GenAI assistant for compliance questions at CISOteria (2024–2025): a pilot with four customer organizations, running on self-hosted open-weights models because the customer documents could not leave their environment.

Background

I have been building production systems since 1993. From 2019 to 2026 I was the backend architect for CISOteria, a security governance and compliance (GRC) SaaS platform, which also powers the Israel National Cyber Directorate’s GRC portal (grc.cyber.gov.il). I owned API and data-layer design, legacy modernization, containerization, SSO / SAML, and the monitoring and audit trails underneath.

Before that I founded a startup (Flixel, MEMS displays, seven patents). I am a Major (ret.) in the Israel Defense Forces and hold a B.Sc. in mathematics and physics from Tel Aviv University.

I run my own infrastructure: bare-metal servers, self-hosted mail, monitoring, and a GPU server for open-weights models. I use what I build.

How I work

Build it, don’t just talk about it. Every system I recommend is one I have built and run myself.

Controls in code, not in the prompt. A prompt can be ignored or manipulated. A permission check, an approval step, or a spend limit in code cannot.

Measure, don’t guess. Which model to use, which threshold to set, whether an LLM is needed at all — each one is a test on real examples. Sometimes the answer is a small fine-tuned model that is faster, cheaper and more accurate.

Fail closed. When an agent hits a limit or an error, it stops and asks a person. It does not keep going and hope.

Simple first, then scale. Start with one task, prove the agent does it well, then give it the next one.

What I’m looking for

Projects where an AI agent takes over real work: customer requests, back-office processes, internal tools, recurring reports. I take on short fixed-price projects, ongoing or fractional-architect engagements, and reviews of agent systems you already run. Best fit: businesses where a wrong action has a real cost — financial services, lending, and any company that handles sensitive records.

I also write about controlling AI spend and abuse at smallestbusiness.com.

Get in Touch

More from me

Controlling the AI you can’t fully trust — its cost, its abuse, and its compliance.

I write near-daily about the practical side — cutting cloud bills and keeping AI spend and abuse under control — at smallestbusiness.com . Get the posts by email, or read them there.

Or read at smallestbusiness.com →