Skip to content
View Jeneidi's full-sized avatar

Block or report Jeneidi

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
Jeneidi/README.md

Mohammad Jeneidi

I build LLM and agent systems, then build the evaluation harness that proves they actually work. Honors CS with a Math minor at Florida State University, 3.88 GPA, class of 2028.

Looking for an AI/ML or applied-LLM internship for Summer 2027. Tallahassee FL, open to remote, hybrid or relocation. (Updated September 2026.)

Selected work

Every one of these ships a measurement, not a demo video.

Project What it does Measured result
lodestone Hybrid BM25 + dense + reranking retrieval engine built from first principles, with the evaluation lab that proves it 96.3% top-5 retrieval on a 300-document SQuAD corpus
toolproof · live Deterministic LLM tool-calling reliability harness. No LLM judges, so the grade is reproducible 40 adversarial cases, 23 tool schemas, 7 failure categories
sidetrack Diffs what a coding agent actually changed against what you asked for, and flags out-of-scope edits before you accept 1.000 precision and recall on a 41-scenario hand-labeled benchmark
mcp-sentinel Stdlib-only scanner auditing MCP server configs for 14 threat classes including tool poisoning and rug pulls 0.867 precision, 1.000 recall over 60 labeled cases
chunklab · live RAG chunking benchmark: 8 strategies scored on 800 SQuAD questions under BM25 retrieval best strategy 93.1% recall@10 against 87.3% for the worst
fetchladder Starts at plain HTTP and climbs to a real browser only once it can prove the cheap answer was a lie. Zero runtime dependencies 7 failure detectors, throws instead of returning a silent zero

Also: driftwatch (live) data and concept drift over 6,587 time-ordered sessions, trueodds (live) probability calibration with temperature scaling and isotonic regression, equipoise (live) fairness audit with an EEOC four-fifths check, promptarmor (live) prompt-injection resistance over a 206-attack corpus, modelroute (live) LLM cost router.

What I am working on now

Differentially private graph neural networks as an FSU research assistant: implementing DP defenses in PyTorch Geometric and measuring where the privacy-utility curve actually breaks. My matrix-decomposition + DP defense is open as PyGIP#35 against the lab's published library. The open question I keep hitting is how much utility a defense can keep once the attacker also gets the graph structure, not just the node features.

Experience

  • AI/ML Engineer Intern, Access to Arabia (May 2026 to present). Azure OpenAI RAG pipeline behind a FastAPI backend, with RBAC, per-document permissions, input validation and secure API key handling.
  • Research Assistant, Florida State University (January 2026 to present). Differentially private GNN methods, privacy-utility evaluation, contributor to the PyGIP library.
  • Software Engineer Intern, Layouti Design (May to August 2025). Internal systems, debugging and testing.

1st in division and 3rd overall out of 130+ teams at the FSU ACM Spring 2026 Programming Contest. Dean's List, Spring 2026.

Stack

Python and C++ first. PyTorch, PyTorch Geometric, scikit-learn, NumPy, Pandas, FastAPI, Azure OpenAI, SQL, Git. Working areas: RAG and hybrid search, embeddings and rerankers, LLM evaluation, prompt and context engineering, agent tooling, model monitoring, differential privacy.

Contact

mjeneidi.com · LinkedIn · Substack

Pinned Loading

  1. fetchladder fetchladder Public

    Start with plain HTTP, climb to a real browser only when you can prove the cheap answer was a lie. Zero-dependency failure detection for web-reading agents.

    TypeScript 1

  2. chunklab chunklab Public

    RAG chunking benchmark: 8 strategies scored on 800 SQuAD questions, BM25 retrieval, best strategy hits 93.1% recall@10 vs 87.3% worst

    JavaScript

  3. lodestone lodestone Public

    Hybrid BM25, dense and reranking retrieval engine built from first principles, with the evaluation lab that proves it: 96.3% top-5 retrieval on a 300-document SQuAD corpus.

    Python

  4. mcp-sentinel mcp-sentinel Public

    Stdlib-only Python scanner that audits MCP server configs and tool definitions for 14 threat classes (tool poisoning, prompt injection, rug pulls); 0.867 precision / 1.000 recall over 60 labeled cases

    Python

  5. sidetrack sidetrack Public

    Diffs what a coding agent actually changed against what you asked for and flags out-of-scope edits before you accept the diff; 1.000 precision/recall on a 41-scenario benchmark, stdlib-only Python

    Python

  6. toolproof toolproof Public

    Deterministic LLM tool-calling reliability harness: 40 adversarial cases, 23 tool schemas, 7 failure categories, zero LLM judges, BYOK live mode for OpenAI and Anthropic

    JavaScript