Eric Kouassi · Senior Full Stack & AI Engineer
AI systems that hold up in production.
Engineering since 2011, applied AI since 2019. GraphRAG platforms, multi-agent orchestration, and computer vision pipelines, with the distributed backends and evaluation systems underneath them.
Open to new opportunities
Reference architecture
A reference architecture for an AI system. An agent runtime sits at the centre and orchestrates six interconnected domains: models covering language and vision, agents, retrieval over vector and graph stores, tools and actions, evaluation including rubrics and red-team testing, and the data platform beneath them.
12+
Years engineering
5M+
Documents indexed
1,000+
Users on production AI
99.9%
Service uptime
Selected Systems
Systems, not screenshots
Each of these is presented as an architecture: the problem it solves and the shape of the solution.
GraphRAG Platform
Humanloop, 2023 to 2025
A production retrieval platform whose index is a knowledge graph, so answers can follow relationships between entities across multiple reasoning steps.
Problem
Vector similarity alone loses the structure connecting facts. Questions spanning several related entities return fragments that look relevant and reason badly, and no amount of reranking recovers a relationship the index never stored.
Result
Supported more than 1,000 active users, with multi-agent orchestration reducing manual analytical workload by roughly half.
- Python
- FastAPI
- Neo4j
- LangChain
- Vector databases
Pipeline
- Ingest
- Extract
- Graph
- Retrieve
- Synthesise
Distributed Semantic Search
Oxide AI, 2021 to 2023
Semantic search and retrieval-based question answering across an enterprise document corpus, with the distributed data pipelines needed to keep it current.
Processed more than 5 million documents on 100 GB to 500 GB pipelines, improving accuracy and performance by 20 to 35 percent.
- Ingest
- Embed
- Index
- Rank
- Answer
Medical Vision Pipeline
viso.ai, 2019 to 2021
Real-time inference over medical imaging, including CT, ECG, and retinal data, built for clinical and industrial deployment.
Reached 92 to 95 percent accuracy while cutting GPU inference latency from roughly two seconds to under 300 milliseconds.
- Capture
- Preprocess
- Infer
- Validate
Product Work
Shipped products
Commercial products I have worked on, live and in production.
- (opens in a new tab)
eesel AI
AI customer support
RAG pipeline and knowledge ingestion
Autonomous AI agents that plug into existing helpdesk and chat tools, learn from a company knowledge base, and resolve support tickets before they reach a human.
- (opens in a new tab)
Macha
AI customer support
Backend AI services
An agent platform for support teams that resolves tickets across a connected stack, reading from helpdesks and acting through commerce and payment tools.
- (opens in a new tab)
SparrowDesk
AI customer support
Full-stack: AI backend and interface
An AI-assisted helpdesk that auto-resolves repetitive enquiries across email, chat, and messaging channels, routing the rest to human agents.
- (opens in a new tab)
Track-POD
Logistics
Independent client project. Sole engineer, from stack setup to deployment.
A last-mile delivery platform covering route planning, live GPS tracking, and electronic proof of delivery for couriers and distribution fleets.
- (opens in a new tab)
Bharat Bazaar
Ecommerce
Independent client project. Sole engineer, from stack setup to deployment.
An online Indian grocery and household goods store serving customers across Japan, spanning several hundred products and nationwide fulfilment.
Expertise
Where the depth is
Grouped by engineering domain, because capability is a question of scope, not of how many technologies fit on a list.
AI Systems
Retrieval and reasoning architectures taken to production.
- LLM applicationsRAGGraphRAGSemantic searchVector databasesPrompt engineering
Agentic AI
Tool-using systems with bounded, observable control flow.
- Multi-agent orchestrationLangChainTool contractsPlanning loopsOpenAI APIs
Evaluation & Safety
Making model output measurable, so improving it is an engineering task and not an argument.
- Golden test setsEvaluation rubricsHallucination detectionPrompt injection defenceRed-team testingHuman-in-the-loop review
Backend & Systems
Service design and the distributed plumbing underneath it.
- PythonFastAPINode.jsTypeScriptMicroservicesREST APIs
Data Engineering
Moving, storing, and retrieving data at working scale.
- Spark / PySparkETL pipelinesPostgreSQLNeo4jRedisMongoDB
Computer Vision
Inference pipelines where latency and false negatives both cost.
- PyTorchTensorFlowOpenCVMedical imagingGPU inference
Cloud & Infrastructure
Getting systems deployed, observable, and recoverable.
- AWSDockerKubernetesCI/CDObservability
Frontend
Interfaces that make complex systems usable.
- TypeScriptReactVue.jsAngularUI architecture
Experience
Working down the stack since 2011
Frontend, then backend, then data and vision, and now applied LLM systems. The range is why the architecture decisions hold up.
2023 to 2025
Senior AI / Full Stack Engineer · Humanloop
Built and ran a production GraphRAG platform, from graph modelling and retrieval through the agent orchestration and evaluation frameworks around it.
- Python
- FastAPI
- Neo4j
- LangChain
- TypeScript
- AWS
2021 to 2023
ML Engineer · Oxide AI
Distributed semantic search and NLP over enterprise-scale document collections, with the data engineering underneath it.
- Python
- Spark
- PostgreSQL
- Vector databases
- AWS
2019 to 2021
Computer Vision Engineer · viso.ai
Enterprise computer vision for healthcare and industrial use, where inference latency and false negatives both carry real cost.
- Python
- PyTorch
- OpenCV
- GPU inference
- AWS
2017 to 2018
Software / Full Stack Engineer · OverApp
Enterprise web applications for finance and insurance, plus the data pipelines behind them.
- Vue.js
- Django
- Node.js
- REST APIs
Approach
How I build
Architecture before implementation
The expensive decisions are made before the first line of code. Getting boundaries right early is what keeps a system changeable two years later.
AI output has to be measurable
A model response that cannot be evaluated cannot be improved or trusted. Evaluation belongs in the deployment path, not in a notebook after the fact.
Reliability matters as much as intelligence
Users do not experience model quality in isolation. They experience latency, failure modes, and recovery, so those are product decisions, not operational afterthoughts.
Simple systems beat clever ones
The cheapest architecture to operate is usually the one with the fewest moving parts. Complexity should be paid for by a requirement, not by preference.
Detailed engineering write-ups are in progress.
Each selected system will get a full breakdown: the constraints, the architecture decisions and their trade-offs, and what the result actually did in production.
Contact
Have a system that needs building?
Open to conversations about production AI systems, platform architecture, and senior engineering work.