Hi, I'm Guy.
Backend & Full Stack Engineer.
3 years building Python backend services and full-stack tools for a production platform. On the side, I got tired of keeping seven sports tabs open, so I built Signal Sports, a system that reads them for me and decides which stories matter. Imagine what I'd build for your team.
Python · FastAPI · React · MongoDB · SQL · Docker · Kubernetes
Signal Sports
Sports news is noisy. A Maccabi Tel Aviv contract negotiation can matter more to one reader than a Champions League final, and a regular-season NBA game can be essential to one fan and pure noise to another. So I built a system that answers one question: what sports news is worth this reader's attention right now?
A background worker pulls five Hebrew sports sites every five minutes, works out what each article is about, groups reports of the same story into one card, and then ranks every story separately for each user. Each user can get a different decision for the same story, along with the reason for it.
My own feed, built from the live corpus. The public demo is a read-only snapshot of it, republished every hour. The UI is in Hebrew.
One article, two readers
These are real headlines, scored by the production engine for two profiles: Guy, a basketball power user who follows Maccabi Tel Aviv, the EuroLeague, the EuroCup and the whole NBA, and a casual fan who follows only Deni Avdija. In one 48-hour window the sources published 480 articles. Guy saw 118 of them, the Deni fan 25.
| Headline | Extracted facts | Guy | Deni fan |
|---|---|---|---|
| Without the injured Deni Avdija, Portland suffer a brutal loss to Minnesotaבלי דני אבדיה הפצוע: פורטלנד ספגה הפסד אכזרי מול מינסוטה | basketball · NBA · Deni Avdija, Portland, Minnesota · injury | high_feed | high_feed |
| Maccabi Tel Aviv lose to Olympiacosבעיטה מפוארת בדלי: מכבי תל אביב הפסידה לאולימפיאקוס | basketball · EuroLeague · Maccabi TLV, Olympiacos · result | high_feed | hidden |
| Detroit stop San Antonio, the Knicks celebrate in Torontoדטרויט עצרה את סן אנטוניו, הניקס חגגו בטורונטו | basketball · NBA · four clubs · result | feed | hidden |
| How Hapoel Jerusalem became the best team in Israeli basketballכך הפועל ירושלים הפכה לקבוצה הטובה ביותר בכדורסל הישראלי | basketball · Israeli league · Hapoel Jerusalem | low_feed | hidden |
The pipeline
The API and the worker run as separate processes sharing one SQLite database in WAL mode. Each five-minute cycle runs this pipeline:
Engineering decisions
01Facts are separate from decisions
Each article is classified once, independently of any user. Relevance is then computed per user from those stored facts, so changing how ranking works never requires re-classifying the corpus.
02Rules first, LLM second, guardrails always
A deterministic classifier handles most articles. The LLM is called only when the rules are unsure. Its output must fit closed enums, and its claims must be backed by evidence in the text. Timeouts and bad JSON fall back to the rules result, and every fallback is counted.
03Abstain rather than guess
A wrong fact reaches the feed as truth, while a missing fact only costs recall, and recall is measured. When the evidence isn't there, the classifier stores "unknown" and records why.
04Safe background processing
The worker takes a single-flight lease before each cycle, so restarts or a second worker can't run overlapping cycles. A failing source degrades the cycle instead of failing it. Notifications go through an outbox and are planned per story, so five outlets reporting the same news produce one push.
05Explainable clustering, no embeddings
Reports are merged through validated story anchors (subject, transaction verb, event state) and Hebrew transliteration matching. Every merge can be traced to one anchor, and a shared name alone is never enough.
06Quality is measured, not assumed
Unit tests show the engine does what it was told. Hand-rated census sets show whether it was told the right thing. Precision and recall are tracked separately, and the census includes a sealed held-out half. Classifier changes are rerun across the whole corpus in memory and diffed before they ship.
Known limitation. Recall is weakest on stories that name no team or player in the taxonomy. In one census, 71 of 76 wrongly hidden stories had no competition attached. That is the next thing I'm working on.
Stack
Want the full walkthrough?
I'm happy to walk through the architecture, the trade-offs and the code on a short call, or you can explore it yourself.
Where I've worked
Full-Stack Developer · IDF Intelligence Directorate, Unit 9900
- As a founding member of the platform team, co-designed and built a shared Python platform that runs 6+ algorithms from multiple teams.
- Cut the time to onboard an algorithm, or adapt it after a major update, from days to hours by standardizing execution, monitoring and output delivery.
- Implemented MongoDB-backed workflow orchestration with status-based locking to prevent duplicate processing.
- Owned the platform's management and monitoring app end to end (React, TypeScript, FastAPI, MongoDB).
- Built GitLab CI/CD pipelines that deploy Docker images to OpenShift, and implemented Kubernetes autoscaling.
IDF Programming Course
Six-month intensive course: object-oriented programming in C#, React and TypeScript, MongoDB.
Let's build something together
I'm looking for my next Backend or Full Stack role in Israel. Whether you have a position in mind, a question about Signal Sports, or just want to talk shop, I'd love to hear from you.
