Simdenis

Systems

Risk Systems — Configurable Data Monitoring + Alerting Web App

Replaced a hardcoded polling script with a configurable monitoring app for datasets, checks, and alerts.

FlaskSQLMonitoring

Built a Flask API and simple web UI for schedules, checks, and recipients.
Integrated MySQL + Snowflake connectors to run user-defined queries.
Added stateful alerting + email notifications with Dockerized deployment.

Research

MIT FutureTech / CSAIL — Science Job Classification + Task Taxonomy

Pipeline to label science jobs, classify research areas, and build a deduped hierarchy of tasks.

LLMsTaxonomyPipelines

Heuristic filtering + LLM classification for OpenAlex/OECD categories with consistency checks.
Prompted task extraction and dedup via embeddings + FAISS.
Clustered tasks with UMAP/HDBSCAN and built hierarchy using MNLI entailment.

Web

Jalapeño Poppers Bot — MIT Dining Menu Alerts

Full-stack menu alert system with magic-link login, keyword matching, and daily digests.

FlaskPostgresEmail

MIT-only magic-link auth with allowlist and rate limits.
Daily menu scraping with caching and robust keyword matching.
Automated digests + unsubscribe tokens; deployed via Fly.io + Neon.

Demo Repo

Web

Oliva UROP — Frontend Prototype for Research Tool

React + Vite scaffold to unblock a research UI before backend APIs were finalized.

ReactViteMUI

Built core UI flows with routing and setup docs for teammates.
Planned integration path for Google Drive/Sheets as an early storage layer.
Focused on fast iteration before backend stabilization.

Repo

Research

LO-Leakage-Bench — Linguistics Olympiad Benchmark + Leakage Tests

Reproducible benchmark runner with isomorphic variants to measure memorization gaps.

LLM EvalBenchmarksReproducibility

Designed a JSONL task format with auto-grading evaluator CLI.
Built an isomorph generator to create surface-form variants.
Reported memorization gaps by task type and difficulty.

NYT Connections — Steering Semantic Embeddings

Research on geometry edits that make embeddings produce more human-like groupings.

EmbeddingsClusteringEvaluation

Created labeled dataset + splits and compared MiniLM vs MPNet.
Tested projection, prototype, and contrastive steering methods.
Evaluated with ARI and accuracy, analyzing difficulty-based failures.

Denis Siminiuc

Notes on me

Tools I use

Selected work

Risk Systems — Configurable Data Monitoring + Alerting Web App

MIT FutureTech / CSAIL — Science Job Classification + Task Taxonomy

Jalapeño Poppers Bot — MIT Dining Menu Alerts

Oliva UROP — Frontend Prototype for Research Tool

LO-Leakage-Bench — Linguistics Olympiad Benchmark + Leakage Tests

NYT Connections — Steering Semantic Embeddings

Notes and essays

The comfort of uncomfortable art

poezie?

Veni, vidi, cucurri: on running and liking it

Kanty Ethics

My Brilliant Review

When will we pass the Bechdel test?

Weekend in New York

Romanian Presidential Elections of 2024; Take Two

writing 123?

Credentials

Connections

Let's talk