Elmir Abdullaiev

Mechanisms and Tools for Data Protection on Web Applications

L. Myrutenko, I. Parkhomenko, E. Abdullaiev, S. Denichenko · PCSITS-2024, Kyiv · April 2024

Detection tools reaching 87% accuracy across 150+ test applications, with OWASP ZAP integration.

Mathematical Modeling of Surfaces Formed by Rotating a Rhombus

E. Abdullaiev · "Shevchenkivska Vesna – 2019", Kyiv · April 2019

Mathematical model and simulations; third place in a city-wide modeling competition.

Agent Fire Drill, a crash test for AI agents

Built with a teammate for the Daytona HackSprint with Arize AI. A mock payment times out after it was processed, and the drill checks the real ledger to see whether the AI agent refunds twice. Across 32 runs on Amazon Bedrock (16 per model), Claude Haiku 4.5 double-refunded 8 of 8 times without a guardrail and 0 of 8 with it, and Claude Sonnet 5.5 stayed safe in all 16. I built the verifier, guardrail and demo. Mock payments, one fault. Interactive demo.

StreetScript, 1st place at AI Agents Hack Night

Team project from the Mistral AI x Elasticsearch hack night in New York. It writes poems and letters using only words cut from 1940s NYC storefront photographs, and a strict check rejects any word with no photograph behind it. I built the Rearrange feature, a hybrid word search and Mistral moderation.

Advocate, a live interpreter that flags what was not a promise

Built in one day at a Google DeepMind hackathon in New York. A renter and a landlord talk through one button: it works out who spoke from the language, translates both ways, and marks the landlord's vague lines, promises and money terms, then writes a two-language summary only after the renter approves. On 20 scripted lines it reached 0.90 precision and 1.00 recall. Not legal advice. Scripted demo.

PharOS, street-safety map and near-miss candidates

Team project from the Real-Time Video Agents Hack in New York. A 3D Manhattan map of a year of crash reports, DOT cameras and live street-video analysis. I built the near-miss measurement (approximate distance and time to collision, shown as candidates for human review) and the shared map color scale. Accuracy was not measured.

Hermes, personal RAG assistant

Personal Telegram assistant, deployed for a single user, with long-term memory (RAG over a SQLite vector store), automatic fallback across six model tiers, Google Calendar via OAuth 2.0, systemd-supervised on a Linux cloud instance.

Rental Housing Law Navigator

Built solo in one night for the Hack-Nation RealPage challenge. An LLM extracts rules from a 59-document public corpus, a verifier rejects any quote not literally in the source, and a deterministic engine decides which rules apply to each of 500 addresses in nine cities. Not legal advice. Live demo.

Chronicle, self-healing project memory

Built the resolution engine for a 4-person team at a MongoDB hackathon: conflict resolution by source authority and recency, and policy learning from human corrections. Verified end to end on real MongoDB Atlas Vector Search.

Security Overflow, OWASP scanner

Flask scanner orchestrating Wapiti, SQLMap and Nikto behind one interface, with LLM-generated remediation guidance. 87% detection accuracy across 150+ test applications.

Utility Debt Shield

Prototype API that flags utility activation friction before a lease is signed. Mock data only, consent-gated, 44 tests with CI.

GenAI guidelines for the university (TUM)

Machine Learning and Society course, July 2024. Systematic evaluation of ChatGPT across 215 academic tasks (87% success rate, 26% false-information rate); findings contributed to TUM's institutional AI strategy.