01
Applied AI &
LLM Systems
Production workflows with explicit control around model behavior.
Forward Deployed EngineerSenior AI Engineer
EmmanuelOnwubuya
I turn ambiguous business problems into production AI systems, data platforms, and products.
Forward Deployed Engineer and Senior AI Engineer working across customer discovery, system architecture, LLM applications, data infrastructure, APIs, deployment, and post-launch iteration.
01datasetevents.delta
02pipelinelakeflow.job
03vector storehybrid.index
04retrieverscope + rank
05LLMroute + tools
06APIFastAPI
Illustrative architecture · no live data
01 / About
Work the problem,
then build the system.
I work at the intersection of AI engineering, data platforms, and product delivery.
My work usually starts with an unclear business problem: understand the workflow, identify where AI or data can actually help, design the system, build the first working version, put it in front of users, measure it, and iterate.
My background in distributed data engineering means I am comfortable below the LLM layer too—pipelines, schemas, retrieval infrastructure, APIs, databases, cloud systems, observability, and production reliability.
I have done this inside enterprise environments, consulting engagements, and products I have built myself.
Operating rangeStakeholder context → system design → production ownership
02 / Expertise
Capabilities across the stack.
I connect customer context to the data, software, and AI infrastructure required to ship.
Applied AI &
02
Data &
AI Platforms
Reliable foundations for retrieval, analytics, and intelligent products.
03
Production
Engineering
The backend and platform discipline needed after the prototype works.
04
Forward Deployed
Engineering
Technical delivery anchored in the workflow and the people using it.
03 / Experience
Built close to the problem.
Client discovery, governed AI systems, and production delivery across healthcare, aerospace, and consulting.
Senior Data & AI Engineer
Frankfurt, Germany- Led discovery with external clients, analysts, and internal teams to uncover reporting needs, access constraints, and workflow pain points; owned delivery from requirements workshops and architecture through demos, user validation, rollout, and production iteration.
- Built client-facing RAG and natural-language analytics over pharmaceutical and HCP data, combining role-aware access, hybrid/vector retrieval, semantic business definitions, and tool calling for governed self-service insights.
- Replaced single-model routing with a policy-based LLM gateway across hosted and self-hosted models, cutting token cost by 12%. Used LangSmith traces to optimize caching, retrieval, and model selection, reducing end-to-end AI response latency by 15%.
- Re-engineered 200+ Snowflake/dbt models on Azure Databricks and Delta Live Tables with CDC and incremental processing, reducing runtime by 50% and platform costs by 5–10% while powering downstream AI workloads.
RAG · LLM gateways · LangSmith · Databricks · Delta Live Tables · dbt
Founder & Principal AI Engineer
Client delivery- Lead client discovery and technical scoping, turning ambiguous business problems into proof-of-concepts, architecture decisions, and deployable AI, data, and software solutions.
- Build LLM-powered applications with tool calling and smart routing through an LLM gateway, optimizing model selection, caching, and execution paths to reduce token cost and latency by 10%.
- Own engagements through deployment and post-launch iteration, validating solutions with users, resolving production issues, and adapting workflows as client data and operational needs evolve.
Discovery · Technical scoping · LLM gateways · Tool calling · Production support
Data Engineer
Hamburg, Germany- Worked directly with Airbus engineering and business teams to discover data needs, refine requirements, and iteratively deliver data products using PySpark, SQL, Palantir Foundry, and Skywise.
- Built and evolved enterprise-scale pipelines and transformation workflows across operational and engineering datasets as client requirements and downstream use cases changed.
- Reduced distributed processing time by 40% through partitioning, predicate pushdown, caching, and query optimization.
- Reduced cloud infrastructure costs by 20% through workload optimization and more efficient resource utilization across Azure and GCP.
PySpark · SQL · Palantir Foundry · Skywise · Azure · GCP
Junior Analytics Engineer
Munich, Germany- Worked with acquisition and finance teams to understand reporting needs and turn fragmented property, acquisition, and financial data into usable analytics workflows.
- Built automated Python and SQL pipelines and API integrations that improved data availability and reduced manual preparation across real-estate reporting processes.
- Integrated data from multiple business systems into centralized reporting, giving teams a more consistent view of portfolio and acquisition performance.
Python · SQL · API integrations · Financial reporting
04 / Selected work
Systems, not demos.
Two products that show how I approach infrastructure, AI behavior, and real-world ownership.
CASE STUDY / 01
Daygent ↗
Static lineage and impact analysis for modern data + AI repositories.
Daygent analyzes source code locally and builds a dependency graph across modern data and AI stacks, helping engineers understand lineage and blast radius before making changes.
It statically understands Python, SQL, dbt, FastAPI, LangGraph, LLMs, embeddings, vector stores, PySpark, Databricks DLT / Lakeflow, Django, SQLAlchemy, and external HTTP dependencies.
Local-first by designNo production credentials. No warehouse connection. No telemetry.
01Repository
↓02Static parsers
↓03Dependency graph
↓04Lineage / impact queries
↓05Offline HTML viewer
$ pip install daygent
$ daygent scan .
✓ dependency graph written
$ daygent graph --html --open
$ daygent impact <node>3,000+users
iOS + Androidin production
Brain Dump converts unstructured thoughts into structured, actionable tasks and scheduling decisions.
Ginja demonstrates end-to-end product ownership: mobile experience, backend services, AI workflows, data, subscriptions, integrations, analytics, deployment, and continued operation for real users.
React Native→FastAPI→LangGraph / OpenAI→Supabase / PostgreSQL
Problem
Turn a stream of unstructured thoughts into a useful plan without forcing users to organize first.
Architecture
React Native clients, FastAPI services, deterministic LangGraph workflows, OpenAI, and PostgreSQL/Supabase.
Production challenges
Reliable structured outputs, notifications, calendar integrations, RevenueCat subscriptions, analytics, and deployment.
Outcome
A live product used by 3,000+ people across iOS and Android—not a standalone LLM demonstration.
System design / How I think
From messy problem to production system.
AI delivery is a loop: the system gets better when discovery and measurement stay connected.
- 01
Discover
Understand user and business workflow.
- 02
Model
Define data contracts and system boundaries.
- 03
Prototype
Validate the smallest useful AI workflow.
- 04
Productionize
APIs, permissions, observability, fallbacks, evaluations.
- 05
Measure
Latency, quality, cost, and usage.
- 06
Iterate
Improve with real-world feedback.
05 / Open source
Tools for understanding complex systems.
I build developer tooling around the hard parts of modern data and AI infrastructure: dependencies, lineage, impact, and change.
MIT License
FEATURED PROJECT
Daygent
Local-first static lineage and blast-radius analysis for repositories that span application code, data platforms, and AI systems.
$ pip install daygent06 / Contact
Building an AI product or trying to turn a difficult workflow into a production system?
I'm interested in Forward Deployed, Applied AI, and senior AI engineering problems where software needs to work in the real world.