DevRel Engineering · Forward Deployed Engineering
Emmanuel
Acheampong
I work the last mile of AI, where frontier research becomes production.
Founding member of the Managed AI DevRel team at Crusoe, working forward deployed: embedded with engineers, building alongside them, turning managed inference into systems that power frontier models. Formerly co-founder and Head of AI at yShade.ai (Google for Startups Black Founders Accelerator '24). Recipient of the O-1 and EB-1A (Extraordinary Ability) visas.
▶ PRESS START
Founder instincts, engineer's hands
I started out building models, not talking about them. As co-founder and Head of AI at yShade.ai, a Google for Startups Black Founders Accelerator '24 company, I led development of computer vision algorithms trained on a 12M+ image dataset, built to work across more than 100 skin shades. Shipping inclusive vision systems into e-commerce and beauty taught me that the hardest part of AI isn't the model; it's everything around it.
That's what pulled me into developer relations. I'm a founding member of the DevRel team for Managed AI at Crusoe, where my job is closing the distance between infrastructure and the open-source ecosystem: writing the integration myself, benchmarking it under real load, shipping the fix upstream, then generalizing it so the next team never hits the same wall.
My work now sits in the unglamorous parts of inference: cache discipline, tail latency, token economics, and the harness design that decides whether an agent is viable at scale. Open models, agentic systems, and the infrastructure underneath them.
- Focus
- Forward Deployed Engineering · DevRel Engineering · Developer Experiences · AI Engineering
- Now
- Senior Manager, Developer Relations, Crusoe
- Previously
- Co-founder & Head of AI, yShade.ai
- Focus
- Open models, agent harnesses, inference infrastructure
- Elsewhere
- pypi/eacheampong
Forward-deployed, then generalized
Developer experience doesn't get designed in a doc. It gets found by sitting with the team that's stuck, fixing the thing, and making the fix permanent.
Embed
Work alongside the teams actually deploying: agent builders, ML platform teams, research groups. Their friction is the roadmap.
Build
Write the integration or reference implementation in their stack, not a slide about it. LiteLLM, LangChain, Vercel AI SDK, MLflow, Pipecat.
Prove
Benchmark under real load and publish the numbers (tokens, latency, dollars) with the methodology attached so anyone can rerun it.
Generalize
Push it upstream and back into the SDK, docs, and cookbook so it stops being a one-off and becomes the default path.
The pattern I keep finding: most inference problems aren't model problems. They're harness problems: a flaky terminal, an uncached prompt, a missing provider entry in a gateway someone's whole platform routes through. Forward-deployed work is how you find those. Open source is how you make the fix outlive the engagement.
What I build
Integrations, agents, and SDKs I've built, mostly making Crusoe Managed Inference reachable from the tools developers already use.
crusoe-litellm-gateway
Your own OpenAI-compatible AI gateway to Crusoe Managed Inference: LiteLLM proxy, Docker Compose, virtual keys, and fallbacks.
langchain-crusoe
A native LangChain integration for Crusoe Managed Inference: the ChatCrusoe chat model with streaming, async, tool calling, and structured output. On PyPI.
crusoe-ai-provider
A Crusoe Cloud managed inference provider for the Vercel AI SDK, drop-in for TypeScript and Next.js agent apps.
mlflow-crusoe
An MLflow deployment plugin for Crusoe's managed inference API: route MLflow deployments straight at Crusoe endpoints.
crusoe-voice-agent
A browser voice agent on Crusoe Managed Inference: Pipecat with local Whisper STT and local Kokoro TTS. Only one API key needed.
deep-repo-agent
Chat with any GitHub repository: a long-context codebase agent powered by Crusoe Managed Inference.
crusoe-huggingface-demos
Hugging Face Spaces demos showcasing Crusoe managed inference: LLM game battles, model arenas, voice, and agent demos.
crusoe-python-sdk
The Crusoe Python SDK: typed access to Crusoe Cloud, with the ergonomics developers expect from a first-party client.
Upstream contributions
- Portkey AI Gateway: added Crusoe Managed Inference as a provider in the open-source LLM gateway
- LiteLLM: first-class provider support, so any LiteLLM-based tool can route to Crusoe models
Things I've written
Developer walkthroughs and benchmarks, mostly about what determines cost and latency once agents start calling tools.
The open agent stack: running Nous Research's Hermes Agent on Crusoe Managed Inference
For agents that fan out parallel tool calls and run unattended, cache discipline and tail latency matter more than they do for a chatbot. A ten-minute walkthrough using two environment variables.
Frontier agents at 10x lower cost: NVIDIA Nemotron 3 Ultra + LangChain Deep Agents
Nemotron 3 Ultra in a Deep Agents harness tuned for it hits top open-model accuracy at roughly a tenth of closed-model cost, and the tuning was all harness engineering, not fine-tuning.
Inference in Production: Engineering LLM Serving for Latency, Throughput, and Reliability
Inference looks simple from the outside: send a prompt, get a response. In production it's a systems engineering problem: KV cache pressure, autoscaling lag, batching tradeoffs, head-of-line blocking, tail latency.
No Single Model to Rule Them All: Building Resilient AI Agents Across Open & Closed LLMs
No one model excels at every task and no single API guarantees uptime. Architecture decisions, failure modes, and performance tradeoffs of moving from a single-model prototype to a resilient multi-model agent.
Building Computer Vision AI Algorithms for 100 Skin Shades
How we gathered the data, trained the models, and deployed computer vision for the beauty and fashion industry across more than 100 skin shades, building for inclusion from the dataset up.
Recent & upcoming venues: WeAreDevelopers World Congress North America (San Jose, Sept 2026) · AI DevSummit + DeveloperWeek Management (South San Francisco, May 2026) · AI DevWorld (Santa Clara, 2023)
What I'm digging into
Neocloud
The new generation of AI-native clouds: purpose-built infrastructure, vertically integrated stacks, and what they change about how AI gets built and priced.
Agents
Harnesses, tool-call loops, and token economics: the scaffolding around a model usually decides whether an agent is viable at scale.
Open models
Frontier-class open weights, the harness engineering that closes the gap with closed models, and the economics that make them win.
Inference
KV cache pressure, batching tradeoffs, tail latency, and cost per token: the systems engineering that decides what production AI actually costs.
Physical AI
Models that touch the real world: robotics, world models, and vision-critical autonomous systems.
Computer vision
The thread from yShade.ai: datasets, evaluation, and vision systems that hold up across lighting, atmospheric conditions, and real deployments.
Financial AI
Models and agents in markets and financial systems: risk, forecasting, and the reliability bar production finance demands.
Get in touch
Happy to talk inference, integrations, agents, or developer experience, and always open to speaking invitations.