Status: In progress
ai · 2024 — present
AI Automation Systems
LLM-driven workflows that remove repetitive work.
- Large language models
- Prompt engineering
- Agent workflows
- +2
Loading
Artificial intelligence
Method
Every model step declares a schema. If the response does not match, it fails loudly instead of flowing downstream.
Nothing takes an action on the strength of a model response alone. Validation sits between the two, always.
If a value can be looked up, it is looked up. Models are for judgement, not for recalling facts.
A workflow you cannot inspect is a workflow you cannot trust — and trust is the whole reason to automate.
Experiments
Decomposing repetitive tasks into constrained model steps with explicit output schemas and validation before any action is taken.
Multi-step agents that call tools and APIs instead of guessing, with every run logged so failures can be traced.
Systematic prompt design — role framing, few-shot examples, output constraints — tested against real inputs rather than cherry-picked ones.
Connecting model steps to real systems so a workflow ends in an action, not a paragraph of text.
Concept: on-device computer vision to distinguish survivors from debris and assess whether a path is passable.
Concept: grounding model answers in a controlled document set so outputs can be checked against a source.
Projects