Enias Cailliau

I help developers & entrepreneurs build with GPT | Founding Engineer at Steamship

Turned down Facebook to build winning startups. 

- Build what customers want.

Co-founder & CTO

The power of compounding.

Architect Reliable AI Agents on a Solid Data Foundation

AI Agents are only as smart as the data they access. I bridge the gap between Data Software  Engineering and Autonomous AI.

👉 The Problem:

Most AI Agent initiatives fail because of poor data infrastructure. Without real-time pipelines and clean context, agents are just toys.

👉 The Solution:

I provide the Data and Context Engineering rigor required for high-performing AI Agents.

What You Gain:

- Smart Context: Engineering the data pipelines that feed your agents the right information at the right time.
- Reliable Execution: Implementing "Human-in-the-loop" and evaluation frameworks to ensure your agents act safely.
- Scalable Ops: From single agents to complex swarms, I architect the cloud infrastructure (AWS/Azure/GCP) to handle the load.

Legacy Integration: Connect your new AI workforce to your existing ERP, CRM, and internal tools seamlessly.
Why Me?
I combine the agility of a startup founder with the architectural discipline of a Fortune 500 consultant.

👉 Let's engineer agents that actually work: [https://topmate.io/raphael_mansuy]

Raphaël MANSUY

OpenClaw-RL: Train Any Agent Simply by Talking

What if your AI agent could get better every time you talked to it — without anyone labeling a single piece of data?

Every interaction produces a "next thing that happens" after the agent responds: you correct it, a terminal spits an error, a test fails. That next-state signal often says exactly how the agent did and how it should improve. Today, deployed agents mostly throw that signal away.

👉 WHY this matters

Think of a student who gets a graded exam back, glances at the score, and tosses it — never reading the teacher's comments. Current RL systems do the same: they might use outcome rewards (pass/fail) but ignore the rich detail in what happens next. A user saying "you should have checked the file first" gives token-level direction. A compiler trace pinpoints the mistake. This is free supervision produced during normal use, and it's wasted.

👉 WHAT OpenClaw-RL does

OpenClaw-RL recovers two types of signals from every interaction:

- Evaluative signals: Did the action work? A re-query implies dissatisfaction; a passing test implies success. These become scalar process rewards via a PRM judge.

- Directive signals: How should the action change? Corrections and error traces carry specific guidance. These become token-level supervision through Hindsight-Guided On-Policy Distillation (OPD).

Key insight: personal conversations, terminal commands, GUI clicks, software tasks, and tool calls are all interactions producing next-state signals. One policy can learn from them simultaneously.

👉 HOW it works

The system runs four fully decoupled async loops — policy serving, environment interaction, reward judging, and training — with zero blocking. The agent keeps serving users while learning in the background.

For OPD specifically:

1. A judge extracts a concise hint from the next-state signal.
2. That hint is appended to the original prompt, creating an "enhanced teacher" context.
3. The same model under this enriched context produces a different token distribution.
4. The per-token gap between teacher and student distributions becomes directional advantage — some tokens are reinforced, others suppressed.

This differs fundamentally from scalar rewards that nudge all tokens the same way. Evaluative (process) rewards and directive (OPD) signals are complementary: in personal agent simulations, combining binary RL with OPD raised personalization scores from 0.17 to 0.81 in just 16 update steps. For general agents, adding process rewards to outcomes improved tool-call performance from 0.17 to 0.30.

The most elegant part: for personal agents, the environment is the user's device. The model improves automatically through normal use — no annotation pipeline, no separate data collection phase.

What strikes me most is the reframing: we’ve treated training as something that happens before deployment. OpenClaw-RL suggests deployment IS training.

Data Engineering | DataScience | AI & Innovation | Author | Follow me for deep dives on AI & data-engineering

OpenClaw-RL: Train Any Agent Simply by Talking

Why I rebuilt a PDF-to-Markdown converter in Rust — and made it zero-config

Every time I needed to extract structured text from a PDF, I ended up in the same painful loop: install poppler, wrangle DYLD_LIBRARY_PATH, pray the native library loads, watch it fail on CI.

The Python tools worked — but they were slow, fragile, and required a runtime environment I didn't want to ship.

So I built edgequake-pdf2md — a Rust CLI and library that converts PDFs to clean Markdown using Vision LLMs (GPT-4.1, Claude, Gemini, Ollama).

What makes it different:

🔋 Self-contained binary — pdfium (~5 MB) is compiled into the binary by default. No runtime downloads. No env vars. cargo install edgequake-pdf2md and you're done.

⚡ Fast — concurrent page processing. A 50-page document costs ~$0.02 with GPT-4.1-nano.

🎯 Accurate — a 10-rule post-processing pipeline strips hallucinated image tags, fixes broken tables, and normalises whitespace.

🌍 Truly cross-platform — macOS arm64/x64, Linux x64/aarch64, Windows x64. All verified in CI.

🦀 Library + CLI — use it from the command line or embed it in your Rust application with a clean async API.

👉 Under the hood:

The bundled feature embeds pdfium at compile time. If you're building from source without a prebuilt copy, build.rs auto-downloads the correct binary for your platform and caches it in ~/.cargo/pdfium-bundle/. Zero friction.

👉 Try it in 60 seconds:

Or via Homebrew:

Inspired by the excellent Python project pyzerox, rebuilt for speed and reliability.

The full source, docs, and CI setup are on GitHub — link in the comments.

Would love feedback from anyone working with PDF extraction at scale. 👇

(Coming soon in egdquake)

taplio

The Battle

Enias Cailliau

Raphaël MANSUY

Followers

Posts

Engagements

Likes

Performance ScoreRaphaël MANSUY's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Enias Cailliau

Raphaël MANSUY

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Raphaël MANSUY's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreRaphaël MANSUY's LinkedIn Performance