I build agent systems that have to be right, not just fluent — and I'm drawn to the part most demos skip: what the thing actually costs, what it silently breaks, and how you would know either way.

Agent & skill engineering. Multi-agent workflows, skill and prompt design, and the measurement layer around them — cost per trigger, context a model can never reach, and honest labels on every number.

Biomedical evidence tagging. Volunteer AI research advisor to a rare-disease research effort (CCHS / PHOX2B): controlled-vocabulary tagging, ontology anchoring to MeSH/MONDO, and grounding-enforced extraction built on one rule — the LLM extracts, deterministic code decides. Same corpus, same tags, byte-identical output across runs and across models.

Hebrew / RTL document pipelines. Bidi-correct PDF, DOCX and PPTX generation. A narrow niche that is genuinely hard and very often done badly.

token-efficient-skill-optimizer — audits AI skills, system prompts, and agent instruction sets for what they actually cost per trigger, under a hard constraint: no task-success loss, no safety weakening.

The design choice I care about most is that it can tell you not to optimize. Several rules exist only to stop an edit — keep the repetition that looks redundant but is load-bearing, keep the verbose instruction carrying a safety obligation. "Already efficient" is a successful outcome, and the pilot that came back at −0.7% was published at −0.7%.

Every number carries an enforced label — [measured], [estimated], [projected] — and a [measured] claim with no pointer to its data file fails the build, including in the project's own reports. 38 rules over 73 sources, each citation machine-checked, and the documentation's own totals regenerated from the registry in CI so they cannot drift. MIT.

visual-video-summarizer — turns a lecture or demo into source-linked study notes with timestamped frames, in Hebrew or English. A deterministic controller owns the pipeline: it names which artifact to write next and refuses to advance past one that is missing, empty, invalid or stale, so a half-finished run fails loudly instead of rendering a confident, wrong summary. Cloud transcription is never implicit — it requires an explicit flag. 407 tests on a Linux/macOS/Windows CI matrix. MIT.

talkbrief — a slide-by-slide brief of a talk where every quote is mechanically checked back against the transcript before it can appear. RTL-correct Hebrew output. 55 tests, CI, typed. MIT.

Based in Israel · Hebrew and English · LinkedIn

Most of what I build is applied and client- or research-facing, so a fair amount of it lives outside this profile. Happy to walk through any of it.