AI / POC Lab
Not Just Aware of AI. Using It.
This section documents what I've actually built, working systems applied to real problems. Not demos for the sake of demos. Each project started with a genuine need, got built across multiple sessions, and is either in active use or in progress. The stack varies. The approach is the same: figure out what AI can actually do by building something that has to work.
Active projects
GPS
gabephilipson.com - This Site
Fully OperationalA production consulting website built entirely with Claude - strategy, copy, prototypes, and full-stack deployment - in under 22 hours at zero build cost.
Last updated: March 29, 2026
Most senior executives either pay $15K-$30K and wait 6-10 weeks for a professional site, or settle for a generic template that undersells them. The goal was a founder-forward dual-audience site with real content and a live AI/POC Lab - without the agency timeline or cost.
What was built
Two-phase build. Phase 1 (Claude chat project): full content strategy, audience positioning, copywriting for all 7 pages, 4 anonymized case studies, AI/POC Lab entries, SEO metadata, brand-consistent HTML prototypes, and a 6-sheet content tracker. Phase 2 (Claude Code, 3 sessions): converted 7 HTML files into a Next.js 15 site, built 10+ shared components, integrated Sanity CMS, wired Resend for contact form delivery, configured Vercel deployment, and connected a custom domain.
What it demonstrates
End-to-end AI-assisted product development - from strategy and copy through technical build and deployment. One person, two AI tools, zero agency involvement. A custom production-grade site that would typically require a 6-10 week agency engagement at $15,000-$30,000 was delivered in under 22 hours at effectively zero build cost.
Screenshots
ADK
An AI-powered college discovery platform built for families. Personalized probability scoring across ~2,000 schools, YouTube-derived student perspectives, and balanced list recommendations.
Last updated: May 7, 2026
Most families don't know how to build a balanced college list. They navigate with spreadsheets, gut feel, and conflicting advice from counselors. AdmitKit asks the right questions, calculates personalized admission probabilities, surfaces student perspectives from YouTube, and recommends a list balanced across Safety, Target, and Reach schools. The tool that started as a custom build for one family is now a public beta available to every family.
What it does
A 7-step survey captures academic profile, intended major, regional preferences, financial generosity needs, and risk tolerance. A probability engine scores every school in a database of approximately 2,000 four-year institutions, accounting for test policy, generosity tier, holistic factors, and academic fit. An AI suggestion layer recommends balanced school sets across Safety, Target, and Reach tiers using a three-layer architecture: deterministic ranker, AI synthesis, and transparency UI. Student Perspective accordion surfaces verified quotes from YouTube videos analyzed across 9 qualitative dimensions. Cost and Financial Aid breakdown shows real net prices and generosity grades.
How it was built
Built solo with Claude Code on Next.js, Supabase, and Anthropic's Sonnet 4.6 model. A custom Common Data Set scraping pipeline extracted GPA percentiles and admission data from 206 of 227 top schools using Firecrawl plus a hardened multi-strategy URL resolver. The probability engine uses selectivity-banded holistic factors and a dual-scenario calculation that accounts for test-optional self-selection. The Student Perspective layer pulls from YouTube using a verified-quote pipeline (96.2% quote accuracy, zero fabrication across pilot runs). Pre-launch security review closed 13 of 16 findings across 5 PR clusters before opening to public beta. 803 tests passing across 37 files.
Screenshots
Search ~2,000 four-year institutions. Each result shows a personalized admission probability and tier (Safety, Target, Reach), with toggles for test-policy fit, regional preference, generosity tier, and major. The full school database is open from the start, not gated behind a survey.
JSA
Autonomous Job Search Agent
Fully OperationalA fully autonomous Claude agent that manages an active VP-level job search end to end -- daily Gmail triage, web-indexed ATS search across 6 platforms, quick-filter screening, tracker management, and follow-up discipline, all triggered by a custom /scan slash command or a weekday cron schedule.
Last updated: April 6, 2026
An active VP-level job search generates more moving parts than most project management systems are built to handle. Dozens of open applications across multiple platforms. Recruiter messages, networking contacts, referral threads. Follow-up deadlines that shift daily. Screening criteria that vary by role, industry, and comp band. ATS portals that don't surface roles in email alerts. Without a system holding all of this together, context gets lost between sessions. Good roles get missed. Follow-ups slip. The search slows down at exactly the point it needs to speed up.
What it does
The agent runs as a Cowork plugin with a custom /scan slash command that triggers the full workflow with zero preamble -- no clarifying questions, no recap, just execution. It also runs automatically at 8:02 AM ET on weekdays via cron. The /scan command accepts natural language variants ("run the scan," "check my job alerts," "any new jobs," "daily scan") and fires the same 7-step workflow every time.
How it is built
The /scan command is defined as a custom skill mapping to a 237-line SKILL.md containing the full execution spec -- every step, every rule, every edge case. CLAUDE.md holds working context between sessions. A memory directory stores the candidate profile, dynamic glossary, and project details. The glossary controls search sources, industry priorities, exclusion rules, and screening thresholds -- all editable without touching code. The Excel tracker is read and written programmatically via openpyxl. A family of related scheduled tasks (job-search-apply, job-search-review, job-search-update, job-search-save) extends the system for specific sub-workflows. The scan skill includes an evals directory for quality testing.
The 7-step workflow
- 1Loads contextReads the candidate profile, dynamic glossary, and Job Search Tracker spreadsheet to establish full session awareness before any action is taken.
- 2Scans GmailSearches for LinkedIn and Indeed alerts, ATS status updates, recruiter responses, and rejection notices.
- 3Searches ATS platformsRuns indexed web queries across Workday, Greenhouse, Lever, SmartRecruiters, iCIMS, and Nigel Frank to catch roles that bypass email alerts entirely. Results are flagged "Verify still open" when posting date cannot be confirmed.
- 4Triages new rolesRuns two-pass deduplication against the tracker, applies quick filters (level, location, function, industry, comp), and classifies each role by AI relevance: core AI role, AI-component, or AI-mentioned. AI-component roles are parked in a holding status rather than closed outright.
- 5Checks follow-upsReviews every open application row by row, flags overdue follow-ups and upcoming deadlines. A 5-day rule flags "To Apply" roles that have aged out. A 10-business-day rule surfaces applications with no response.
- 6Drafts summary emailCompiles a structured HTML report with clickable job description links and drafts it directly to Gmail.
- 7Updates the trackerWrites new rows, status changes, and closures to Job_Search_Tracker_MASTER.xlsx via openpyxl, and timestamps the update.
Screenshots
+2 more
BGB
BuyGiveBack — Mission-Driven Ecommerce
In ProgressA full-stack ecommerce platform built with Claude Code that donates 50% of net profit to NYC school districts.
Last updated: April 6, 2026
Discount retail has no differentiation. This platform ties every purchase to a transparent, publicly auditable giving formula. Customers see exactly how their savings translate to donations for underserved NYC school districts.
What it does
Sells below-retail products and donates 50% of net profit to underserved NYC school districts. The donation formula is public-facing and calculated transparently. At checkout, customers see their cart savings and are prompted to donate that amount. A public dashboard shows live totals by school district.
Tech stack
Next.js 15 App Router on Vercel, Supabase for database and auth, Stripe for payments, Resend for transactional email. Entire codebase built using Claude Code from initial scaffold through feature implementation.
Screenshots
▣
Add screenshots in Sanity Studio
Upload images via /studio to populate
TII
A satirical 2026 Mets wins predictor built on a riff on CAPM. Live record, tunable intangibles, hardcoded fWAR.
Traditional baseball projections like PECOTA are statistically rigorous and emotionally empty. Fans know a season is not just wins above replacement. It is tabloid drama, mascot charisma, and whether the bullpen can find the strike zone in August. This project puts those variables in the equation.
What it does
Calculates projected 2026 Mets wins by summing a replacement-level floor, fWAR contributions, a tunable intangibles term, and an error bar. Current-season record, streak, and 162-game pace pull live from the free MLB Stats API. Preset buttons let visitors flip between optimistic, pessimistic, and historical scenarios.
How it was built
One HTML file with embedded CSS, JS, and base64 logo and mascot assets. A Vercel serverless function proxies the MLB Stats API with 24-hour edge caching so the public API gets hit at most once per day per edge region. The page falls back to hardcoded values silently if the API fails.
Screenshots
▣
Add screenshots in Sanity Studio
Upload images via /studio to populate
CMP
Your personal AI product manager. Describe a problem in plain English, get a prioritized strategy, and ship engineering-ready requirements with stakeholder-tailored documentation. 60 features live across 9 categories. Seven pipeline runs across software, creative, and business domains.
Last updated: April 8, 2026
Phase I complete. Phase II active. Stakeholder intelligence shapes agent decisions and tailors BRD/PRD exports to individual audiences. Seven documented runs across ERP forecasting, e-commerce, athletic training, furniture making, and professional networking, with 100% schema compliance since Round 2. The pipeline adapts naturally to non-software domains without special handling.
What it does
Takes a PM from raw problem statement to approved PRD through a structured multi-agent pipeline. Each stage (Discovery, Strategy, Requirements) produces structured outputs the PM can review, edit, and approve before advancing. An Orchestrator agent manages state, routes rejections back to the earliest affected upstream stage, and maintains a running session dossier stored as JSON in GitHub. Gate modes are configurable per stage: manual, supervised, or auto with a countdown timer. Discovery is permanently locked to manual — it is the foundation, and wrong problem means wrong everything. Each stage delivers formatted output with risk and priority badges. Stakeholder context shapes Strategy rationale and Requirements flags. /cmp:export doc generates a six-section BRD/PRD. /cmp:export doc [name] tailors the executive summary to the named stakeholder's communication style and priorities. 49 features live, 11 planned.
How it's built
React single-page app with a provider abstraction layer supporting Claude (default), OpenAI, and Gemini — swappable per agent without touching agent logic. Session state persists as a structured JSON dossier in a private GitHub repo. Each agent reads the full dossier and writes only to its own block. The Orchestrator carries a hybrid context model: structured JSON state plus a running narrative summary. A pre-flight intake form (with optional /cmp:brief free-text command) scores project size before firing the pipeline. Seven documented runs with consistent token efficiency (~17,800 avg). Phase II adds stakeholder intelligence, BRD/PRD export (delivered), open question answer flow, and Execution, Launch, and Analytics agents.
Screenshots
Compass — PM Agent Pipeline
Your personal AI product manager. Describe a problem. Get a strategy. Ship requirements.
Phase II, Active
Phase
II of III
Phase I complete, Phase II active
Features
60
49 live, 11 planned
Pipeline runs
7+
Software, creative, business
Avg tokens/run
17,775
Full pipeline end-to-end
How it works
Discovery
Frames the problem, identifies target user, surfaces assumptions with structured risk classification.
→
Strategy
OKRs, ranked features with rationale, explicit out-of-scope. Picks the right framework for the project.
→
Requirements
User stories with Given/When/Then acceptance criteria, NFRs, open questions with blocking dependencies.
Each stage is reviewed and approved by you before the next one runs. Automation is earned, not assumed.
The Compass hero — Phase II active with 60 features, 7+ pipeline runs, and an average of 17,775 tokens per full run from problem to requirements.
In progress
NC
New Compass
Governed multi-agent product platform · Claude, OpenAI, Gemini
In Progress