01/Index
Writing ↗⌘K
01Index

Zarif Latif

▹Founder @Railo

I build intelligent systems — from research to production.

Get in touch→RésuméGitHubWriting ↗Open to select opportunities
02About

About

I’m a software engineer and AI researcher working at the intersection of artificial intelligence, software engineering, and cybersecurity. My research spans machine learning, language models, mechanistic interpretability, adversarially robust security, reinforcement learning, program analysis, and automated program repair. I’m particularly interested in building intelligent systems that are not only capable, but reliable, secure, and explainable.

I’m the founder and technical builder of Railo, a GitHub App that fixes security findings in pull requests with deterministic AST-level rewrites, gated by independent checks, and of LIFE-FORGE, an open-source harness that red-teams LLM agents. I enjoy taking difficult problems from first principles and turning them into working systems, and I measure my tools against real code and publish the results, including the corrections.

03Selected Work

Selected Work

A selection of systems I've designed, shipped, and researched.

01
DevSecOps & Program Analysis

Railo — Autonomous DevSecOps & Security Remediation

A GitHub App that fixes security findings in pull requests with deterministic AST rewrites (no LLM writes the patch). Every fix is checked by a rescan, a structural check for the known-safe construct, and the project's own tests; only fixes that clear every check can auto-merge, and the rest go to human review. Opened upstream: security fixes in Dagster and BentoML (path traversal) and Plane (client-IP validation), plus bug fixes in HTTPie and DeepEval — all open for maintainer review, none merged yet.

Live ↗
5 Upstream PRs · In Review
2025 – Present
  • Python
  • LibCST
  • Semgrep
  • TypeScript
  • FastMCP
  • Docker
  • PostgreSQL
02
AI Safety & Agentic Simulation

Life Forge — Autonomous Flight Simulator for AI Agents

An autonomous adversarial red-teaming simulator for AI agents: 3D MAP-Elites quality-diversity evolution breeds edge-case scenarios that force frontier models to fail before production. In a three-model showdown it discovered 7 unauthorized fund-transfer executions in Qwen 2.5 and unbounded retry deadlocks in Llama 3.1 and DeepSeek-R1. MCP-native sandbox with deterministic rollback and an invariant policy oracle.

Code ↗
89/89 Tests · CI Green
2026
  • Python 3.10+
  • 3D MAP-Elites
  • Model Context Protocol (MCP)
  • LiteLLM
  • FastAPI
  • Docker
  • Pytest
03
Quantitative Risk & NLP

Aetherius — Deterministic Downside-Risk Engine

Open-source quantitative risk intelligence pipeline that screens concentrated public-equity books against real-time SEC EDGAR filings and global financial news feeds using deterministic scoring taxonomies. Backtested with 100% recall across SVB-2023, Wirecard-2020, and FTX-2022 with a median lead time of 2.48 days and 0 false positives.

Code ↗Paper ↗
100% Recall · 2.48d Lead Time
2026
  • Python 3.12
  • SEC EDGAR API
  • GDELT DOC 2.0
  • FastAPI
  • Pytest (Hypothesis)
  • Docker
04
AI Engineering & Multi-Agent

SignalForge — Generative Engine Optimization (GEO)

An open-source Generative Engine Optimization (GEO) and AI Share-of-Voice measurement framework. Measures how AI answer engines (ChatGPT, Claude, Perplexity) rank and cite B2B brands, and deploys autonomous multi-platform earned-media squads to seed LLM training and RAG retrieval pipelines.

Code ↗
Multi-Engine GEO Measurement
2025
  • Python 3.11
  • FastAPI
  • React (Vite)
  • PostgreSQL
  • Llama 3.1
  • Ollama
  • Groq
  • Docker
04Research

Research

Peer-reviewed and in-progress work in machine learning.

My research centers on Mechanistic Interpretability, Adversarial Robustness & AI Safety, Automated Program Repair (APR), AST Program Synthesis & Transformation, and neurosymbolic ai systems.

2026Under Review · ACM TOSEM (CORE A*)

VeriPatch: A Verification-Gated Neuro-Symbolic Framework for Secure Program Repair

Md. Zarif Latif, Kazi Tasfin Mahmud

ACM Transactions on Software Engineering and Methodology (TOSEM)

★ ACM Artifact Badges Candidate

2026Working Paper · Target: KBS (Q1)

Prescriptive Graph Bandits: Safe Causal Policy Learning via Counterfactual Graph Neural Networks and Distributional Offline Reinforcement Learning

Md. Zarif Latif

Knowledge-Based Systems (Elsevier, JCR Q1)

2026Working Paper · Ready for Submission

Lightweight Explainable Static Malware Detection with Temporal Generalization and Adversarial Robustness

Md. Zarif Latif, Tasneem Tuhfa

Journal of Information Security and Applications (Elsevier JISA)

2026Working Paper · In Preparation for TACL

The Tokenization Ceiling: Why Induction Heads Fail for Bengali Despite Being Language-Agnostic at the Token Level

Md. Zarif Latif

Transactions of the Association for Computational Linguistics (TACL)

05Stack

Stack

AI & Machine Learning

  • PyTorch & Transformers
  • LLMs & Alignment (RLHF)
  • Mechanistic Interpretability
  • Adversarial Robustness
  • Model Evaluation & Red-Teaming
  • Sparse Autoencoders (SAEs)
  • RAG & Neural Retrieval

Security & Program Analysis

  • Automated Program Repair (APR)
  • AST Analysis & Rewriting
  • Static Analysis & SAST
  • Autonomous DevSecOps
  • Vulnerability Remediation
  • Adversarial AI Defense

Languages

  • Python
  • TypeScript
  • Rust
  • C / C++
  • Go
  • SQL
  • Bash / Shell

Systems & Infrastructure

  • Distributed Systems
  • Docker & Kubernetes
  • Next.js / React & Tailwind
  • FastAPI & Microservices
  • PostgreSQL & Redis
  • Linux Systems & Toolchains
  • CI/CD & Cloud (AWS / GCP)
06Contact

Contact

Let's build something
worth doing.