AI & Machine Learning Company in Tel Aviv

Tel Aviv ML work comes in three sizes: a scoped AI proof of concept at Tel Aviv rates, a custom production ML model (fraud scoring, claims triage, churn prediction, document extraction, defect detection on a vision line), or full production AI systems with RAG, multiple models, fine-tuning, on-prem or il-central-1 inference, computer-vision components, and enterprise integrations. Scope drivers include a data audit, a baseline model, a hosted demo, and an initial Israel AI Policy 2023 risk classification. Tel Aviv rates sit above most European hubs because of the AI21, Mobileye, and Run:ai talent premium and the 8200 alumni pool, but well below Bay Area benchmarks for comparable senior teams. We give fixed-fee proposals in shekels with a USD reference, not open T and M estimates.

The Israel AI Policy, published in 2023 by the Ministry of Innovation, Science and Technology and the Innovation Authority, is the country's principles-based framework for responsible AI. It does not yet impose binding obligations the way the EU AI Act does, but it sets expectations on transparency, accountability, fairness, safety, and human oversight, and it signals the direction the PPA, Bank of Israel, and Ministry of Health will follow in sector-specific guidance. Our discovery phase runs an Israel AI Policy classification on your use case (limited risk, high risk, or sensitive sector), and high-risk builds ship with a model card, a bias audit, a dataset lineage report, and a monitoring plan. For clients exporting AI services to the EU, we layer EU AI Act conformity work on top, because the Act applies to non-EU providers placing systems on the EU market regardless of where the model was trained. The PPA has telegraphed that Israeli regulators will not wait for primary legislation before applying enforcement pressure on opaque high-risk systems.

Yes. Hebrew is a low-resource language relative to English, and most frontier closed-source models handle it competently at the surface but degrade on legal, medical, and rabbinic-text use cases. We ship RAG systems, internal copilots, document extraction pipelines, and agent workflows for Israeli banks, insurance carriers, healthcare systems, and government-adjacent buyers, defaulting to AI21 Jamba and Jurassic for Hebrew-native enterprise workloads, fine-tuning Llama 3 or Mistral on client Hebrew corpora when sovereignty or cost matters, and falling back to Anthropic and OpenAI through Azure Israel Central or AWS Bedrock when the use case is bilingual or English-dominant. We apply strict PII redaction tuned for Israeli identifiers (Teudat Zehut national-ID format, Israeli phone numbers, Hebrew name patterns), prompt-injection defences calibrated for Hebrew and Arabic mixed input, human-in-the-loop review gates for any customer-facing output, and evaluation harnesses that run on a held-out Hebrew test set before every production push. Output logs feed existing SIEM and model-risk platforms.

Every engagement begins with a privacy and risk classification that maps to both the Israel Privacy Protection Law 5741-1981 (with the 2017 Data Security Regulations and the recent amendments tightening sensitive-data rules) and the EU AI Act if the client serves European customers. For PPA we deliver a data-protection impact assessment, a register of processing activities, a documented legal basis for each data source, and the Hebrew and English notices the PPA examiner expects on a first inspection. For EU AI Act we map the system to its risk tier (prohibited, high-risk listed in Annex III, limited, or minimal), and for high-risk systems we deliver the technical documentation pack (Annex IV), a quality management system aligned with Article 17, conformity assessment artefacts, and post-market monitoring runbooks. We coordinate with Israeli privacy law firms when interpretation questions arise and do not give legal advice ourselves. For defence-adjacent dual-use work, we layer Defence Export Control review on top before any code or weights leave Israeli soil.

When a project requires genuine research (novel architectures, low-resource Hebrew or Arabic NLP, reinforcement learning in production, sample-efficient learning on small medical datasets, post-quantum-safe inference protocols, novel watermarking for GenAI outputs), we scope collaborations with Technion, Weizmann, Tel Aviv University, Bar-Ilan, or Hebrew University AI groups rather than overselling in-house capability. Israeli academia produced foundational work in reinforcement learning, theoretical machine learning, and computer vision, and the Weizmann group's work on RL and inverse RL is internationally cited. Our core team handles applied engineering, MLOps, fine-tuning, productionisation, and edge deployment, which is where most Tel Aviv AI projects actually stall. For standard work (RAG on existing foundation models, classical tabular ML, computer vision on known architectures like YOLO and DETR, agent orchestration with LangGraph or LlamaIndex) no academic partner is needed. We will tell you up front which bucket your problem fits into, and we have introduced clients to Technion and Weizmann industry-affiliate programs when the roadmap and budget fit.

Since 2023, all three hyperscalers run native Israeli regions. We default to AWS il-central-1 (Tel Aviv) for training, model artefacts, vector stores, feature stores, and inference endpoints. Azure Israel Central and GCP me-west1 (Tel Aviv) cover disaster recovery and cross-cloud high availability. For Ministry of Health clinical AI, Bank of Israel-supervised credit and underwriting models, Ministry of Defence dual-use vision and sensor projects, and any workload touching Teudat Zehut national-ID data, this is the first time a fully sovereign-cloud AI stack is achievable in Israel without standing up your own GPU footprint. We document residency in the architecture diagram, the data-processing record under the Israel Privacy Protection Law 5741-1981, and the model card. Cross-border to AWS Frankfurt or Ireland is acceptable only for non-sensitive analytics and only after a documented impact assessment, and cross-border to US regions requires explicit sign-off because of Defence Export Control implications for vision and autonomy workloads.

A typical Tel Aviv insurance or bank model (fraud scoring, claims triage, credit decisioning, AML transaction monitoring, KYC document extraction) takes fourteen to twenty-two weeks from kickoff to production, assuming clean historical data and PPA and Bank of Israel documentation as part of scope. Week 1 to 4 is data audit, baseline modelling, and Israel AI Policy risk classification. Week 5 to 12 is modelling, iteration, challenger-model testing, and Hebrew-specific evaluation if the model touches text. Week 13 to 18 is MLOps on AWS il-central-1, observability wiring through Aporia or Arize, shadow-mode deployment behind the existing production model, and internal audit review against the carrier's model-risk policy. Week 19 onward is gradual rollout with population-segment monitoring and explicit bias-audit gates. If you are pre-data or need labelling (common for Hebrew document-extraction projects), add six to eight weeks for an annotation phase using Israeli annotators rather than offshore generalist labelling services.

Israeli AI research and development frequently qualifies for Innovation Authority (Rashut Hahadshanut) grants, including the R&D Fund (typically thirty to fifty percent matching on eligible research spend), the Ideation Incentive Program for very early-stage exploration, and the Technological Incubators track for novel applied-research startups. Eligible activity generally includes novel modelling, architecture experimentation, algorithmic uncertainty, Hebrew and Arabic NLP advancement, sample-efficient learning on small datasets, and explainability research, not routine integration of existing APIs. We deliver time-tracked engineering logs, experiment histories, failed-hypothesis documentation, and the technical narratives an Innovation Authority committee actually reads, and we structure projects so the eligible scientific-research portion is clearly separable from production-engineering work that does not qualify. The Trade-Up program supports later-stage international expansion of AI products. We do not draft the grant application itself; that sits with your dedicated Rashut Hahadshanut consultant and the Authority's assessment committee.