Skip to content
View elmir132's full-sized avatar

Highlights

  • Pro

Block or report elmir132

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
elmir132/README.md

Elmir Abdullaiev, AI and software engineer, M.Eng. Computer Science at Cornell Tech. From a 4.62 KB model on a microcontroller to RAG assistants and live voice agents. Opens elmirabd.me.

Hi, I'm Elmir

AI and software engineer, M.Eng. CS candidate at Cornell Tech.
Background in cybersecurity (B.Sc. honors, M.Sc. in progress), now focused on applied ML systems that ship.

📍 New York, NY  ·  🎓 Cornell Tech, M.Eng. CS (May 2027)  ·  elmirabd.me

Selected projects

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

Neural classifier for real-time Bluetooth Low Energy threat detection on embedded hardware. Published at CH&CMiGIN-2026, Kraków (CEUR-WS Vol-4278, Scopus-indexed).

Unsafe runs: Claude Haiku 4.5 8 of 8 without the guardrail and 0 of 8 with it; Claude Sonnet 5.5 0 of 16; 32 runs in total, 16 per model
Python Daytona Amazon Bedrock Arize
99.68 percent accuracy, 4.62 KB model on an ESP32-S3, 75.5 percent smaller after INT8 quantization (from 18.88 KB), per the paper
TensorFlow Lite Micro ESP32-S3 BLE 0.9 ms

Address-level housing-law lookup. A language model proposes rules from the source documents, a verifier rejects any quote that is not literally in the source, and a deterministic engine decides which rules apply. Built solo overnight for the Hack-Nation RealPage challenge. Not legal advice. Live demo.

Team project (with Pranav, Divyesh and Sahil) that writes poems and letters using only words cut from 1940s NYC storefront photos, every word traced back to its photo. I built the Rearrange feature, a hybrid vocabulary search and Mistral moderation.

59 source documents, 500 addresses, 9 cities
Python Streamlit Gemini
1st place, team of 4, 165 tests
Python Elasticsearch Mistral

A renter and a landlord talk through one button. It works out who spoke from the language, translates both ways, and flags vague commitments, promises and money terms. Built in one day at a Google DeepMind hackathon. Not legal advice. Scripted demo.

Telegram assistant deployed for a single user, with long-term memory: RAG over a SQLite vector store, quota-aware routing across several models with automatic fallback, supervised by systemd on a Linux server.

On 20 scripted lines: 0.90 precision and 1.00 recall
Gemini Live FastAPI JavaScript
6 model tiers with automatic fallback, 20 tests
Python SQLite RAG systemd

More projects

Project What it is
Security Overflow Flask scanner that orchestrates Wapiti, SQLMap and Nikto behind one interface, with LLM-generated remediation guidance. Successor to my B.Sc. thesis scanner.
PharOS Team hackathon project (4 people): a 3D Manhattan crash-history map with live street-video review. I built the near-miss measurement and the shared map color scale.
Chronicle Self-healing project memory from a MongoDB hackathon (team of 4). I built the resolution engine: conflict resolution by source authority and recency, and policy learning from human corrections.
Utility Debt Shield Prototype API that flags utility activation friction before a lease is signed (mock data only).
ClipTidy macOS menu-bar app that cleans clipboard text with one shortcut.
Multi-model research orchestrator CLI that compares several AI providers on one question and shows where they disagree.
tuichat Full-screen terminal chat client for Ollama and OpenAI-compatible models.
Crypto trend dashboard, NFT call builder Small Flask and Node tools on public market data.

Stack

Languages: Python, TypeScript and JavaScript, C and C++, SQL. ML and AI: TensorFlow, TensorFlow Lite Micro, PyTorch, Gemini, Amazon Bedrock, RAG, Arize. Backend and data: FastAPI, Flask, MongoDB Atlas, PostgreSQL, SQLite, vector search. Infra: Docker, systemd, Azure, Daytona. Hardware: ESP32-S3, BLE, MQTT. Opens the repository list.

Contact

LinkedIn: in/elmirabd Website: elmirabd.me Email: ea488@cornell.edu

Banner and badges are generated by scripts/build.py. Every number comes from the linked project READMEs.

Pinned Loading

  1. fire-drill-core fire-drill-core Public

    Crash-test AI agents against a real ledger: a payment times out after it was processed, does the agent refund twice? Daytona + Amazon Bedrock + Arize. Interactive demo included (mock payments).

    Python

  2. streetscript streetscript Public

    StreetScript: write with only words cut from 1940s NYC storefront photos, every word traced to its photo. 1st place, AI Agents Hack Night (Mistral AI x Elasticsearch), team of 4.

    Python

  3. advocate advocate Public

    Live interpreter for a renter and a landlord that flags vague commitments, promises and money terms. Gemini Live, FastAPI. Not legal advice.

    Python

  4. hermes hermes Public

    Personal Telegram AI assistant (single user) with RAG long-term memory over SQLite, quota-aware model fallback and Google Calendar, Gmail and Todoist tools.

    Python