Large products where the AI does the repetitive work, reliably.
Production AI automation built into large business products — turning slow, manual, repetitive business processes into LLM-driven pipelines that run unattended, stay observable, and fail safely. The work at Citifuel and across the systems documented here: automation that an operations team can actually trust with real workload.
Most 'AI features' are demos that break the moment they meet real business data, volume, or edge cases. A business process can't be automated by a clever prompt alone — it needs retries, validation, audit trails, human-in-the-loop checkpoints, and a clear answer for what happens when the model is wrong. An automation that's right 95% of the time and silent about the other 5% is a liability, not a feature.
LLM and NLP services wrapped in async job pipelines (FastAPI, Celery, Redis) so the heavy generation never blocks a request. Each automation is a defined process — structured inputs, validated outputs, retries, and an audit record — with humans able to review or override at the points that matter. RAG and prompt pipelines keep the model grounded in the business's own data rather than guessing. The hard part of AI automation isn't the model — it's everything around it: the retries, the guardrails, the trail you can audit afterwards.
Treat the model as an unreliable worker that needs supervision, not a black box to trust. Make every automated step observable and reversible. The goal is not to remove people, but to remove the repetitive work that was never a good use of them — and to do it without quietly introducing errors at scale.