PRD Agent
Building
A small Claude-powered agent that turns a rough feature idea into a structured PRD:
problem statement, user stories, and acceptance criteria. It decides on its own when to
pull in market context, flag compliance risk, or format the final document.
Python
Anthropic API
Tool use
Streamlit
A custom GPT that works like a professional frozen-dessert formulator. Describe a flavor,
such as saffron ice cream or roasted fig and honey gelato, and Batch 83 returns a fully
weighed, ratio-balanced recipe built for the Ninja Creami.
Its formulation logic is anchored in the whole-wheat sorbet base ratios developed by NYC
chef Alex Stupak. Every recipe is calibrated against professional targets for sugar, fat,
solids, and stabilizers, then delivered with a step-by-step method, a diagnostics table
(including PAC and POD), and variations for softer, firmer, or lower-sugar results.
Custom GPT
Prompt engineering
Domain-specific rules
Recipe scaling
Menu Pricing Copilot
Planned
An AI pricing assistant for independent restaurants. It costs out every dish from its
recipe and current ingredient prices, then recommends a menu price based on target
food-cost percentage, plate margin, and local competitor pricing. When supplier prices
move, it flags which dishes have drifted out of margin and suggests a fix.
It applies the revenue-management thinking I used in hospitality pricing to a business
where most operators still price by gut feel and spreadsheets.
Pricing strategy
LLM agents
Unit economics
Margin alerts
Cookbook Knowledge Assistant
Planned
A retrieval-augmented assistant over my personal cookbook library. Ask "how do I stop
gelato from getting icy?" or "what can I make with leftover egg whites?" and it answers
with citations to the exact book and page. A held-out question set scores each answer for
accuracy and groundedness, so retrieval changes are judged by data, not gut feel.
RAG
Vector search
Citations
Eval harness