Eric Krock repository of development projects and software product management work. Available for new opportunities in product management for AI, industrial software, enterprise software, or digital health. Contact me on LinkedIn.
Try out the demo version! ElectroMiser is an agentic AI application that automates storage and export decisions to save money for Tesla Powerwall owners on California’s NEM 3.0 tariff. (1) A tool-calling RAG agent answers natural language user questions about power generation and usage. Claude selects from 12 read-only tools (live Powerwall status, today’s energy curve, 30- and 90-day average solar and load curves, historical period stats, forecasts, rate schedule, etc.), decides which data it needs over several iterations, then synthesizes an answer and sometimes emits an inline chart. A router tiers the model and iteration budget by question complexity. Simple factual questions get Haiku and a 3-iteration cap, while open-ended questions get the full model and budget. Whisper and OpenAI TTS make the same loop voice-driven. (2) An autonomous control agent runs a continuous perceive, decide, act loop to optimize reserve and export against tariff rates. It ingests live telemetry, irradiance forecasts, and utility rate schedules, computes projected minimum state of charge, and writes battery reserve and export-mode changes to the Tesla Owner API unattended. Because it controls hardware, it includes hysteresis to prevent thrashing, exponential-backoff write gating, and a hard floor so a bug can’t drain the battery below a safe reserve. (3) Agents also run my development process: Claude Code writes the code, an Opus LLM judge grades AI-generated user-facing text against an 11-scenario regression harness before release, and a security-review agent audits the codebase. Implemented using Claude Code in Python and hosted on AWS EC2. Uses: pypowerwall library to retrieve data from Tesla Owner API; Anthropic Claude Sonnet — primary RAG generation (Q&A, daily summaries, EV charging advisor); OpenAI Whisper-1 — transcribes voice queries; OpenAI TTS-1 (alloy voice) — speaks the answer back for hands-free use.; InfluxDB Cloud — time-series storage; Open-Meteo — GTI solar forecast and cloud-cover data; ASHRAE photovoltaic production estimation algorithm to estimate expected production over the course of the year (theoretical open skies, and derated for expected monthly cloud coverage).
Built agentic AI system to automate job search and resume/cover letter workflow. System scans Gmail job alerts hourly and target company web sites daily and runs every lead through a filter stack: a company blocklist, target-title match, banned-keyword filter, compensation floor, travel-percent cap, and a location/time zone policy that reasons about intent. Remaining leads are given priority score. Best matches get a fully tailored resume and cover letter, drafted by a subagent that pulls from my own approved wording, checks every claim against a source-of-truth facts file, and gets independently QA-reviewed by an adversarial QA agent. A background watcher runs the whole pipeline continuously, so new leads show up triaged and ranked, not as a raw unsorted inbox. HITL final review. Streamlit dashboard, SQLite database, Claude Code subagents, Microsoft Word editing integration.
Try out the demo version! Built simulation for modeling throughput and performance for a single robotic back end assembly line or multiple lines working together to build electronics products. User can download a CSV file to set up the simulation parameters, upload the edited CSV simulation configuration file, run the simulation, and see the results. Can define types of robots and define each type’s capabilities and vary number of stations, parts to build, target ticks/unit, distance to central store, number of robots of each defined type, and input/output buffer sizes. See Help for details.
I want to keep up with what’s going on in my city. But I don’t have the time to read all of every email from the City Council, City Manager, and every city commission or even to scan every email’s table of contents. Replit made it easy to write an email filtering agent that connects to a dedicated inbox via IMAP, parses plain text, rich text, and PDF attachments, applies rules to identify and summarize City of Sunnyvale content of interest to me, and delivers summary emails with embedded source links. I’d built an earlier version of this with greater difficulty and an inferior result using n8n. Replit gave me a better result in just hours!
Used Lovable and Google Gemini to develop a prototype digital health tracker app for people with Body Focused Repetitive Behaviors such as nail. Created synthetic imagery to validate model. Used Gemini to assess each finger’s health. In the future I plan to create a HIPAA-compliant version suitable for production consumer use. Earned “Build and Deploy an AI App” certification from NextGen Product Manager.
Used n8n, Twilio, and GPT-40-Mini to create a phone-based vaccine information interactive voice agent providing scientifically accurate answers to questions about vaccine safety and efficacy. Curious? Call USA + 833-655-1585 and ask it questions about vaccine safety and efficacy! Earned “Agentic AI: Voice Agent with Workflow Orchestration” certification from NextGen Product Manager.
Used n8n, JavaScript, and Outlook OAuth 2 API to create an agent for filtering emails from city government, analyzing their contents, and forwarding a daily summary of emails of interest to the subscriber. Earned “Agentic AI: Workflow Orchestration” certification from NextGen Product Manager.
Used Chat_GPT 5.2 to write custom GPT that helps you debunk long social media posts (e.g. Gish Gallops on Facebook) that contain multiple false statements. Paste in thee text of a post found on social media. It will identify up to 10 false claims and write a reply calmly debunking each one, citing credible sources. Responses are presented as flat text no longer than 8000 characters to make them easy to copy and paste into a Facebook reply. Wrote validation test harness in Python within Google Colab. Read the FAQ.
Used Chat_GPT 5.2 to write custom GPT that helps you correct online misinformation without escalating arguments. It takes a single false claim and (usually) writes a calm, platform-ready reply (for X, X Community Notes, Bluesky, Facebook, etc.) with credible sources: 240 characters max, neutral, and easy to copy and paste. Wrote validation test harness in Python within Google Colab. Read the FAQ. Earned “Create a Custom GPT” certification from NextGen Product Management.
Personal Life Coach app developed as team project using Lovable and Large Language Model. Wrote UI framework, user profile manager, and family retirement, insurance coverage, college savings financial analytics, and LLM integration to create personalized financial plan. Earned “Shipped an AI Agent” certification from NextGen Product Manager.
Agentic AI solution for managing supply chain risk developed as team project using Lovable, Python, n8n, OpenAI, publicly-available risk-information databases. Wrote Watchman Agent that implemented back-end database and APIs for multi-agent coordination and monitored product delivery commitments, BoMs, and inventory on hand to detect when inventory was running low. Also wrote simulator to simulate day-by-day manufacturing activity. Risk Scanner Agent scanned public databases of risk information to identify when changes in risk environment put supplier commitments at risk. Sourcer Agent evaluated alternative suppliers for their risk and recommended optimal supplier for next shipment. Earned capstone project certification from NextGen Product Manager.