Software Engineer — GenAI / Applied ML
Sakhnin, Israel
Final-year B.Sc. Software Engineering student (Feb 2027) specializing in GenAI and applied ML. Builds production-style backend systems — Airflow pipelines, PostgreSQL, async Python — and runs controlled measurement studies across RAG, video ingestion, LLM serving, and agentic pipelines.
INFRA
100% vs 0%
Showed that throughput tuning which works on an A100 does not preserve latency SLOs on an L4 — preventing a costly wrong hardware-porting decision.
EVAL
$0.11/run · 3.2% FP
Built a model-drift and cost watchdog that re-benchmarks LLM providers nightly, catching silent provider regressions before they reach users.
GENAI
13 pp
Demonstrated that agentic self-correction in RAG helps a strong model (+6.0 pp) and hurts a weaker one (−7.0 pp) — an interaction that decides whether the added complexity is worth paying for.
BACK
2,069 → 424
Cut a 90-day analytics query to ~424 buffer reads (~5×) by matching an index's shape to the query's own access pattern, confirmed with EXPLAIN rather than guessed.
Semantic Video Intelligence & Predictive Reasoning Platform
Dual-stream video embeddings (SigLIP 2 + InternVideo2) with reciprocal-rank fusion, a multimodal RAG stage producing cited reports, and a four-tier Qwen2.5-VL verification cascade. Containerized microservices: React, FastAPI, RabbitMQ, Qdrant, PostgreSQL, MCP server.
Two-person, faculty-supervised. Phase A advisor-approved; Phase B Sep 2026 – Mar 2027, in progress.
React · FastAPI · RabbitMQ · Qdrant · PostgreSQL · MCP
Model Drift & Cost Watchdog
Nightly re-benchmarking of LLM providers against a fixed 40-task suite, with rolling-window z-score drift detection, version-aware baselines, and database-enforced idempotency. 480/480 results at $0.1134 per run; 3.23% measured false-positive rate over 100 simulated noise runs.
Built and verified locally against real providers.
Airflow 3.x · Python 3.12 · asyncio + httpx · PostgreSQL 16 · SQLAlchemy 2.x · Docker Compose · pytest
Satellite Ecological Monitoring & Heat Anomaly Detection
Streamlit platform over NDVI and Land Surface Temperature for 8 ecological sites in Northern Israel, Landsat and MODIS via Google Earth Engine, 2000–2026. Includes a second independently built RAG pipeline (ChromaDB + Gemini 2.5 Flash) over an academic corpus, and z-score heat-anomaly detection at three severity levels.
Two-person, Ecological Models Lab, Spring 2026.
Streamlit · Google Earth Engine · ChromaDB · Gemini 2.5 Flash · scikit-learn
Peer-to-peer academic help platform
Six-person team, 2025. Role: system engineer.
Next.js 14 · Tailwind · Firebase Auth + Firestore · Vercel · three OpenAI-backed API routes
A three-part controlled measurement programme testing one question: does the optimal configuration of an AI pipeline depend on the hardware it runs on? Two of the three hypotheses were disproved and published as such.
96 frozen configurations swept identically on A100 and L4. Pareto membership differed (4 vs 6, 3 shared, Kendall's τ-b = 0.80) — but via the embedding model, not the reranker the hypothesis named.
Under 0.1% drift above the model's 384 px input but 1.4% at 360p — a 13.4× jump, unanimous across 600/600 frames, correcting common 360p guidance to 480p for this model.
Throughput payoff transferred almost exactly (2.56× vs 2.70× at concurrency 128); SLO compliance did not — A100 met the budget at every level, L4 at none.
2×2 factorial on HotpotQA (n = 100): naive vs agentic self-correcting RAG across GPT-4o-mini and Llama-3.1-8B. +6.0 pp on the stronger model, −7.0 pp on the weaker.
SARSA + safety-based shaping reached 71.46% ± 2.74% vs 47.71% ± 5.76% for the best MC variant, over 20 runs × 10,000 episodes with 95% CIs.
Caught a fabricated benchmark result before publication by noticing an on/off ratio of exactly 1.00× across all eight test levels, then added the liveness assertion that made the failure impossible to repeat.
Python · RAG · LangChain · LangGraph · vLLM · FastAPI · PostgreSQL · Airflow · Docker · PyTorch · embeddings · vector databases (Qdrant, FAISS, ChromaDB) · LLM evaluation · asyncio · pytest