ai agent
Kernel
An AI agent that tests a product idea against what customers have already said, and returns a sourced case for and against.
the problem
A brand team with a product idea has to decide whether to spend on it, with customer signal (reviews, search, social mentions, competitor moves) in different places and read reactively, if at all.
why it hurts
Reading it by hand is slow, so most ideas reach a spending decision without being tested against what customers have already said.
what i built
Kernel: an AI agent I built for Slurrp Farm, a $90M consumer foods brand, on one corpus of 15,000-plus customer verbatims. A product manager types a hypothesis and Kernel answers from real customer reviews: a sourced case for and against, a go-to-market plan and a confidence-rated verdict. I pitched it to the founders directly and built it end to end with AI.
how it works
Let code do everything a model should not.
whyRetrieval, counting and checks are code. The writing model sees only a frozen evidence pack, so it cannot invent a customer or a number.
Sort the checks by what they protect.
whyA fabricated quote withholds the whole pack; a banned word warns and still delivers.
Keep the fixed pipeline as the baseline.
whyAn agentic version that chooses its own next step runs beside the fixed pipeline, which is kept as the baseline it is measured against.
in detail
A hypothesis in, sourced customer evidence out, and a call that comes with its conditions.
Kernel
A product manager types a hypothesis, say a single-serve millet cookie pack for school tiffin boxes, and Kernel answers from real customer reviews: a sourced case for and against, a go-to-market plan and a confidence-rated verdict.
The call, from the evidence alone
The writing model sees only the frozen evidence pack and may not add a number, a customer or a quote. In this run, fifteen checks passed before the call was shown.
What would have to be true
The call comes with its conditions, and each assumption it rests on is marked against the evidence, contradicted or still unknown.
why it's useful
The brand team can test an idea against what customers already said before spending on it. Every claim traces to the sentence behind it, and the writing model cannot invent a customer or a number. Kernel was delivered with Mill, its companion tool, in three months, and both are now used by the founder, insights and GTM teams.
technical specs
- Postgres full-text retrieval, then a model extracts closed-taxonomy tags and verbatim excerpts in batches, code counts frequencies and confidence bands, a second model writes from that frozen pack alone, code gates the result, and a person decides.
- An agentic version that picks its own tools runs beside the fixed pipeline, with guardrails against hallucination and evals against that fixed baseline.
- The checks are sorted by what they protect: a fabricated quote withholds the whole pack; a banned word warns and still delivers.


