OHOMER HAYATRésumé ↗

PROJECTS

Projects

Selected projects and current work.

Selected projects

Internship Monitor eligibility-first listing pipelineLive Greenhouse and Lever listings are canonicalized, then a deterministic eligibility gate evaluates role, geography, season, graduation, language, and work authorization. An optional local intelligence layer with Qwen, RAG, and tools supplements the gate without overriding it. Provenance is stored in SQLite before alerts are queued.ELIGIBILITY-FIRST PIPELINEInternship MonitorGREENHOUSElive job boardsATS feedLEVERadditional sources15 sources totalCANONICALLISTINGSnormalized roles2,800+ live listingsDETERMINISTIC GATEhard eligibility · cannot skiprolegeoseasongraduationlanguagework authLOCAL INTELLIGENCEQwen · RAG · toolsoptional · supplements, never overridesSQLITE +PROVENANCEstate + audit trailALERT QUEUEidempotent notifications15 SOURCES2,800+ LISTINGS70-CASE HUMAN REVIEW221+ TESTSBOUNDED RETRIESIDEMPOTENT NOTIFICATIONS

Internship Monitor

An internship-intelligence pipeline that discovers live roles, evaluates likely eligibility, adds optional local LLM intelligence, and safely queues explainable alerts.

2,800+ listings · 15 sources · 70-case human review · 221+ automated tests

Built withPython · AsyncIO · SQLite · Qwen · Ollama · RAG · GitHub Actions
View code ↗
RepoTriage confidence-aware issue intelligenceA GitHub issue is classified into likely labels, sent to human review when confidence is low, matched with similar historical issues, and persisted with maintainer feedback.CONFIDENCE-AWARE ML SYSTEMRepoTriagePANDAS ISSUE #59214bug reportLABEL SCORESBug92%IO81%Docs31%CONFIDENCE GATEpredict or abstainSIMILAR ISSUES#475890.88#441020.83#398760.76TEMPORAL SPLITSCALIBRATED THRESHOLDSTF-IDF RETRIEVALMAINTAINER FEEDBACK → POSTGRES

RepoTriage

A confidence-aware GitHub issue intelligence system that predicts labels, retrieves similar issues, abstains when uncertain, and records maintainer feedback.

Built withPython · FastAPI · scikit-learn · Next.js · TypeScript · PostgreSQL · SQLAlchemy · Docker
View code ↗
FailSpec bug reproduction and regression testing pipelineA bug report enters an isolated Git worktree investigation, Codex forms a reproduction hypothesis and generates a Playwright test, then deterministic verification classifies the sanitized evidence.LOCAL-FIRST BUG REPRODUCTIONFailSpecBUG REPORTReact / Nextstructured report + repo pathHYPOTHESISCodex inspects worktreeforms reproduction planisolated git worktreePLAYWRIGHT TESTtest('repro checkout crash', async () => {…})constrained generated specVERDICTVerifiedpartial evidencenot reproducedexecution errordeterministic checkCODEX ANALYSISISOLATED WORKTREESANITIZED EVIDENCE286 PASSING TESTS

FailSpec

A local-first developer tool that turns bug reports into evidence-backed Playwright regression tests through isolated execution and deterministic verification.

My role
Platform architecture & backend orchestration

286 tests · isolated Git worktrees · controlled Playwright execution · deterministic verification

Built withTypeScript · React · Express.js · Playwright · OpenAI Codex
View code ↗
PitchMirror multimodal presentation-analysis pipelineA presentation video is converted into keyframes and transcript data, processed through an asynchronous AWS pipeline, and returned as structured feedback across voice, presence, and content.MULTIMODAL COACHINGPitchMirrorvideo • keyframes • transcriptS3mediaSTEPorchestrateECSprocessBEDROCKanalyzeCOACHING REPORTVoice84Presence76Content91schema validatedIDEMPOTENT JOBSAUTOMATED REPAIRDETERMINISTIC FALLBACKSASYNC AWS PIPELINE

PitchMirror

A multimodal presentation coach that turns recorded pitches into structured feedback on voice, presence, and content.

My role
Solo build

Thousands of frames → 5 keyframes · Step Functions → ECS Fargate → Bedrock · idempotent execution

Built withNext.js · Fastify · FFmpeg · AWS · Amazon Bedrock · Amazon Nova Pro
View code ↗
Kaboom procedural CPU ray-marched fireballA code-generated fireball rendered with signed distance fields, sphere tracing, procedural noise displacement, finite-difference normals, lighting, and OpenMP-parallelized CPU frame generation.DEPENDENCY-FREE CPU RENDERERKaboomcameraFRAME 084640 × 480120 framesSDF tracingfBM noiseOpenMP CPUSPHERE TRACINGFINITE-DIFFERENCE NORMALSLAMBERT + RIM LIGHTINGPPM → GIF / MP4

Kaboom

A dependency-free CPU ray-marching renderer that generates a procedurally animated fireball using signed distance fields, sphere tracing, and procedural noise.

Built withC++ · OpenMP · Ray Marching · Signed Distance Fields · Procedural Rendering
View code ↗
Tiny Shell Unix command and process pipelineA Unix-style shell parses a command, creates child processes using fork, runs programs using exec, and connects their input and output through a pipe.UNIX PROCESSES FROM THE INSIDETiny Shelltiny-shell$cat notes.txt | grep TODOtokens:catnotes.txt|grepTODOparser → pipeline with 2 commandsPARENT SHELLparse commandpipe() + fork()fork()CHILD Aexec("cat")CHILD Bexec("grep")pipefd[1]fd[0]COMMAND PARSINGFORK + EXECPIPE + DUP2PARENT / CHILD PROCESSESC

Tiny Shell

A Unix-style shell written in C to explore command parsing, process creation, program execution, and pipelines from the inside.

Built withC · Unix · fork/exec · Pipes · Systems Programming
View code ↗

IN PROGRESS

The workshop

MatchLens chronologically evaluated Premier League predictorHistorical Premier League matches are converted into past-only rolling form features, split chronologically, evaluated against naive baselines, and inspected with class and probability diagnostics.FOOTBALL ML + HONEST EVALUATIONMatchLens5-MATCH ROLLING FORMCHRONOLOGICAL SPLITtrain 80%test 20%MATCH OUTCOMEHOME52%DRAW18%AWAY30%PHASE 1 DIAGNOSTICSAccuracy0.5121Macro F10.3640Tests24 passingdraw failure remains visiblePAST-ONLY FEATURESNAIVE BASELINESARTIFACTIZED METRICS + PROBABILITIESLOGISTIC REGRESSION
ACTIVE

MatchLens

A Premier League prediction system focused on leakage-safe features, chronological evaluation, strong baselines, and honest model diagnostics.

Built withPython · pandas · scikit-learn · matplotlib · pytest
View code ↗
+

Whatever catches my attention next.