AI engineering
Models are components. The pipeline is the product.
The AI work I have shipped is automation engineering in the old sense: queues, retries, schemas, and thresholds — with a model doing one job inside a pipeline that does not sleep.
Built for throughput, measured in executions, judged by failure rate.
The numbers come from production runs, not a slide. Counters settle once on first view; the server renders the final values.
5,000+
Executions per day through the Polarions distribution engine
1,000+
Telemetry data points per second on the dementia detection build
35%
Fewer pipeline failures after the mentoring rounds
The shape of the pipeline.
Two models, one router, one orchestrator, one publishing surface. Everything between them has a contract and a failure path.
Distribution engine — Polarions, Sweden
Content that moves between platforms without a human in the loop.
The engine ran on n8n, with the Meta Graph API as the publishing surface and Cloudflare Workers as the edge layer. Mistral and Groq models sat behind OpenRouter: one model drafts, one model rewrites for the platform, and neither sees a prompt that says “be creative tonight”.
My job was mostly the boring parts — queues that drain, retries with ceilings, idempotent writes, and the dashboard that says a healthy day is five thousand runs and a handful of failures. When failure rates rose, the fix was mentoring, not magic: four interns learned to write modular steps, and the pipeline failure rate fell by a third.
n8n · Meta API · Cloudflare Workers · OpenRouter · Mistral · Groq
5,000+
Workflow executions per day, without a human approving each one
−35%
Pipeline failure rate after the modular-design mentoring rounds
4
Interns taken from zero to owning pipeline steps
Telemetry under load — Smart India Hackathon
A dementia detection system that watches data the way it arrives.
Flutter on the device, REST APIs behind it, and a stream of behavioural telemetry flowing at a thousand data points per second. The architecture question was the same one as in fintech: what is the shape of the payload, who owns the schema, and what happens when a second arrives late?
The answering system treated the signal as a pipeline, not a demo: typed payloads, bounded queues, and evaluation that ran offline on recorded sessions before it was trusted anywhere near a person.
Flutter · Dart · REST APIs · behavioural telemetry
1,000+
Data points per second processed through the telemetry APIs
0
Minutes lost to a schema argument — the payload format was decided first
Practice — how the models are held
A model is an intern with excellent pattern match speed.
I treat LLM work like staffing a junior team: one job per model, a contract for its output, a reviewer for anything published. OpenRouter keeps the roster swappable — Mistral, Groq, and whatever is cheaper next quarter, all behind one interface.
The tooling earns its place the same way everywhere else: n8n and Activepieces for orchestration, typed schemas at the boundaries, and a failure budget I can explain to a client. The stack also includes the adjacent skills — RAG pipelines where retrieval matters, structured outputs where parsing would rot.
OpenRouter · Mistral AI · Groq · n8n · Activepieces · RAG
1
Job per model — anything broader than one job gets split
3
Orchestration platforms in production experience: n8n, Activepieces, Zapier
Throughput beats cleverness.
A pipeline that quietly does five thousand correct things a day is worth more than a demo that does one impressive thing. Design for the unexciting Tuesday.
Fallbacks are architecture.
Models go down, inputs arrive mangled, and rate limits arrive mid-launch. The system is the retries, the ceilings, and the dead-letter queue — not the prompt.
Measure the failure rate.
“83% fewer hallucinations” is marketing. A pipeline failure rate, measured before and after an intervention, is engineering. I reduce the one nobody printed on the slide.
The questions that come up on the first call.
Straight answers, before you book anything. If yours is not here, ask it directly — the call is free and there is no pitch deck.
Ask on a free callFull detail — the roles, the stack, the outcomes you can print — lives on the résumé. The fastest conversation is a free call.