smlflg/smlfg

Samuel Fleig

AI Engineering student · building Human-Agent Interfaces

Projects · HAI · Agent context · How I build

Hi, I’m Samuel

I build systems for working with AI agents without giving away human ownership.

My work sits between agent orchestration, context systems, verification loops, and AI-native learning tools. I care less about making agents appear autonomous and more about making human intent executable, inspectable, and recoverable.

• 🎓 Studying AI Engineering
• 🧠 Building with Codex, Claude, Hermes, local tools, and knowledge graphs
• 🛠️ Interested in agent reliability, context engineering, evaluation, and human agency
• 🔭 Turning real friction into small, testable systems

Human-Agent Interface

HAI is my attempt to build a better control layer for working with AI.

Explore the project at human-agent-interface.com.

human intent
    ↓
context · routing · memory
    ↓
scoped agent work
    ↓
verification · handoff
    ↓
owner decision

Agents should increase a person’s capacity to act — not quietly replace their ability to decide.

What I build

System	Focus
HAI / Hermes	Human-owned agent routing, roles, recovery, and handoffs
Knowledge systems	Source-aware context retrieval instead of invented certainty
Agent evaluation	Reproducible checks for prompts, tools, and agent workflows
Learning systems	AI tutors and interactive tools that preserve understanding
Small product prototypes	Turning an observed problem into a useful, verifiable artifact

How I build

observe a real failure
        ↓
define a narrow owner-controlled task
        ↓
build the smallest useful artifact
        ↓
verify against concrete evidence
        ↓
keep or discard the system

• Context before action.
• Small scopes before broad refactors.
• Evidence before confident claims.
• A green tool output is not automatically a finished outcome.
• Human approval matters when a decision has real consequences.

Public build principles

What “human-owned” means

An agent may research, draft, build, test, or summarize.

But it should not silently redefine the goal, overwrite important work, or turn uncertainty into a confident story. Important decisions remain visible and attributable to the owner.

What I mean by verification

A file existing is not proof that a system works.

I prefer concrete checks: real outputs, test runs, source contracts, runtime evidence, and an explicit statement of what was not verified.

System pulse

_{Public GitHub signals only · generated daily from public API data · no private activity or token data}

Explore

• Visit human-agent-interface.com
• Browse my public repositories
• Start with the projects pinned on my profile
• Follow the artifacts: system design, source code, checks, and demos

---

_{Building systems where humans and agents can do more together — while humans remain the authority.}