Curriculum Vitae
Ami Heines
AI Engineer & Backend Architect — LLM systems, evaluation and guardrails
Summary
Backend architect and team lead with thirty years building production systems, now working on LLM applications. In 2024 I built and piloted a retrieval-augmented GenAI assistant over a corpus of security and compliance frameworks, evaluated by CISOs — a domain where an invented control ID is a liability rather than a glitch, which is where I learned that grounding and verification are the engineering, not the model. Since then I have built Guardrail, a fail-closed AI gateway enforcing spend and abuse limits in the request path, and an open-source RAG service on Postgres / pgvector scaled to 35M chunks. For seven years, to 2026, backend architect and team lead for CISOteria, a governance-risk-compliance (GRC) SaaS spun out of IPV Security, which also powers the Israel National Cyber Directorate’s GRC portal — so I am used to building software where a regulator eventually asks what happened and why. Former startup founder (Flixel — MEMS displays, 7 patents), IDF Major (ret.), B.Sc. Mathematics & Physics, Tel-Aviv University.
Selected projects
GenAI assistant for GRC — pilot, CISOteria
2024–2025- — Built and ran an internal RAG assistant letting CISOs ask compliance questions in plain language instead of navigating control trees. Piloted with four tenants and evaluated by CISO domain experts; scoped as a product feature but handed over when the company brought in a specialist AI subcontractor, so it did not reach general release.
- — Through 2025 I was the technical counterpart to that subcontractor: briefed them on the corpus and its structure, on which retrieval behaviour had held up under expert review and which had not, and on what the pilot never covered — tool calling, write actions, retrieval tuning. They rebuilt on their own stack; my part was making sure they inherited the findings rather than repeating the pilot.
- — Ran open-weights models from Hugging Face on a GPU instance in the company’s own AWS account, behind an Open WebUI deployment — no third-party LLM API — because customer compliance material could not leave their environment.
- — Assembled and prepared the corpus: ~45 regulatory frameworks (NIST CSF 2.0, FedRAMP High / Moderate, PCI-DSS v4, ISO 27001 with SOC 2 and IASME mappings, CMMC, C2M2, HITRUST CSF, Cyber Essentials, EU Cyber Resilience Act mapping) plus ~114 customer policy documents anonymized before ingestion — sources where a plausible but invented control ID is a liability, not a glitch.
- — Whole-corpus search or a pinned document to scope the answer, because expert users wanted control over provenance rather than a black box; per-tenant scoping of customer documents through the platform’s existing API permission model.
- — Selected the model by hand-running the same GRC question set across candidates, trading response latency against answer accuracy under a fixed GPU budget — one instance, so model size traded directly against tokens per second.
- — Wrote a custom Open WebUI filter pipeline in Python that augments the conversation with live external data: parses a vendor / product reference out of the user’s message, queries a self-hosted OpenCVE API, filters to CVEs updated in the last seven days, and injects them with an explicit instruction to recommend only from the supplied CVEs and cite which apply.
LangGraph banking copilot — personal build, synthetic banking data
2026- — interrupt()-based human approval gating every money-moving action, with the approval decision and the resume persisted in checkpoint history as an audit record. The control is in the graph topology, not in the prompt — no path reaches a write tool without passing through the approval node.
- — Multi-agent copilot on LangGraph: intent router to retrieval to tool-calling agent to verification node, with checkpointed state. Typed action layer in Pydantic, split into read tools and money-moving write tools; authorisation enforced in code against the authenticated session, never in the prompt.
- — Two-layer eval suite: deterministic control tests that call the tools directly and assert the authorisation and money-conservation invariants — offline, ~1.5s, wired into a pre-commit hook — plus conversational evals in pytest + DeepEval, traced in LangSmith. The control tests are mutation-checked: removing an authorisation check makes exactly those tests fail.
Guardrail — Fail-closed AI gateway
2026- — A fail-closed AI gateway for LLM apps and agents. Today it enforces hard spend caps and denial-of-wallet limits; being extended toward the wider guardrail stack — prompt-injection defense, PII redaction, output validation — integrated from proven tooling rather than rebuilt.
- — Designed against the OWASP LLM Top 10, EU AI Act, and ISO 42001 — enforcement in the request path, not a policy document. Packaged as a LiteLLM plugin.
RAG on Postgres / pgvector — open source
2026- — Retrieval service — chunking, embeddings, similarity search — on self-hosted Postgres with pgvector and a local Ollama inference server, deliberately without a dedicated vector database.
- — Scaled the same design to a 35M-chunk English Wikipedia index with HNSW and ~1 ms queries.
- — Exposed it as a public demo at ask.amiheines.com. The model behind it is small and self-hosted, so the answers are ordinary — the part worth showing is the operational envelope: an endpoint anyone can hit, with layered rate limiting at the edge and a GPU concurrency semaphore and token budget in the application.
CISOteria — Security & Compliance (GRC) SaaS · Backend Architect & Team Lead (contract)
2019–2026- — Backend architect and team lead for a multi-tenant CISO / compliance platform — owning API design, the data model, and overall architecture. Also powers the Israel National Cyber Directorate’s GRC portal (grc.cyber.gov.il), built during my last year there; the platform was being onboarded to Project Nimbus, the Israeli government cloud, when I left.
- — Node.js / Express microservices alongside a large legacy PHP codebase I incrementally modernized; MariaDB, Redis, Docker; NVD / CVE vulnerability data into a management dashboard; SSO / SAML, AWS CloudTrail, Terraform.
- — Led a small distributed team across Israel and India, and guided the automated-test practice.
Smallest Business — Cloud & AI-Cost Engineering
2026–present- — Advise teams on cutting cloud bills and controlling AI / LLM spend without re-architecting the product — request-path cost controls, caching, model routing, egress and compute reduction.
- — Built the supporting tooling in Astro / Node.js; near-daily writing at smallestbusiness.com.
Earlier: NFT pack-opening contract on EOSIO (2022); biometric blockchain wallet (2021); enterprise cyber & privacy application, IPV Security (2019); encrypted-SMS lottery automation, Philippines (2017–2018); automated online lending platform, USA — Lendmart (2016–2017).
Experience
Technical skills
- GenAI / LLM
- Production LLM integration, RAG (chunking, embeddings, hybrid search, grounding), prompt engineering, tool calling & typed action layers, self-hosted open-weights inference; evals and tracing; inference cost & latency optimization — model routing, semantic caching, context right-sizing
- Languages
- Python, JavaScript / Node.js, PHP, SQL, C++
- AI security & governance
- AI gateways & spend/abuse control (denial-of-wallet), LLM guardrail architecture; OWASP LLM Top 10, EU AI Act, ISO 42001, NIST AI RMF
- Backend & architecture
- Node.js / Express, PHP (legacy modernization), REST APIs, microservices, multi-tenant SaaS
- Data & caching
- PostgreSQL (+ pgvector), MariaDB / MySQL, Redis, Elasticsearch
- Infra & DevOps
- Docker, Terraform, AWS, nginx, Linux, CI/CD, SSO / SAML, Prometheus / Grafana / Loki
- Blockchain
- Antelope / EOSIO, BFT consensus, smart contracts (C++ / WebAssembly), private-chain deployment
- Frontend
- Angular, React, D3.js
Education
Recognition & patents
- — EOSIO Stack Exchange — ranked #11 globally (above Daniel Larimer, creator of EOS)
- — 7 international patents in MEMS display technology (1998–2004)
- — SPIE Conference: Bi-Stable flat-panel display (Proceedings of SPIE Vol. 4712, 2002)
- — Commanding Officer, Major (ret.), Israel Defense Forces
- — Member of Mensa International
Available for short- and long-term engagements. Book a scoping call →