About

I’m Ahmed Gamal AKA Snow in X. I’m an AI engineer mostly working on retrieval with RAG, Agents Harnesses and Context Engineering. I’m a Freak when it comes to following best practices (Such the ones from Anthropic’s how to build reliable agents and other technical blogs) to then invent and come up with my own ideas.

It started with Interstellar and ElizaOS by a16z. I wanted to bring TARS to life at midgnight, tinkering about giving it a personality, real data from NASA, and a humor function it could call when I told it to dial itself down (if you know you know). The curiosity behind that idea back when agents were still unknown is exactly what shaped who iam today.

Before AI, I audited smart contracts (Solidity + CodeHawks/Sherlock.xyz), me and the team i was with were already AI native in the sense of using it explore more bugs, as a matter of fact a cool FUN fact is that we’ve already implemented a principle called “PRIMERS”, a principle that everyone in the world are already using, except that it’s now called “SKILLS.md” the idea derived by simply created MD files each specific to a vulnerablity type allowing the agent to learn patterns.

Since when i was young i’ve always had an itch for Research, started off as a Smart Contract secuirty researcher, now an AI Applied Engineer, and soon an AI Researcher engineer.

I’m fluent in Arabic and English, and I’m genuinely excited to be at the frontier of where AI goes next.

Outside of work I box, watch films, play chess, and tinker with code combining ideas.

What I build

Harnesses, not prompts. Most of the work is the loop around the model: planning, tool calls, state, retries, and the context the model actually gets to see. I write the design down before I write the code, because the harness is where reliability is won or lost.

Evidence-first retrieval. Hybrid dense-plus-sparse retrieval fused with reciprocal rank fusion, deterministic guardrails before any model call, and fail-closed citation gates. The system cites the source or refuses. It does not guess.

Deterministic cores, typed edges. LangGraph graphs with Pydantic contracts and secured tools, while plain Python owns routing, money math, and side effects. A model reads messy input and writes prose; tested code decides what happens next, and risky cases wait for a human.

Evaluation as part of the deliverable. Versioned golden sets, Ragas suites, and LLM-as-judge gates where deterministic checks cannot answer the question. The misses get published next to the wins.

Day to day: Python 3.12 and 3.13, LangGraph, Deep Agents and Pydantic AI, FastAPI, Postgres, Qdrant, React, Docker, and Docling, with models on AWS Bedrock, OpenAI, Groq, Cerebras and NVIDIA, and tracing through LangSmith.

Selected work

ProjectWhat it does
OcclusionDental patient FAQ assistant over an openly licensed corpus: pre-LLM guardrail, hybrid retrieval, verified citations or a safe refusal.
NutriMindSix-agent nutrition planner with an LLM-as-judge quality gate and a caloric-safety interrupt.
agentic-support-triageSupervisor-routed support triage with typed outputs, secured tools, and a deterministic plus LLM-judge eval suite.
shipment-exception-triage-agentLogistics exception triage with deterministic tools, evidence logging, and human-in-the-loop approval.
MCP-ProjectPaste a job URL, get a resume-matched gap analysis from a LangGraph agent backed by two MCP servers.

The Occlusion write-up is the longest thing I have published: read it here.

Before this

I spent 2025-2026 auditing Solidity smart contracts for reentrancy and access-control bugs in competitive audits, which is where “find the specific way this breaks” became a habit. Before that, a year in retail sales, which is where I learned to explain hard things to people who do not care about the internals.

Open to AI engineer roles, remote or on-site.

Contact