Paul Chambers

Voice-first PWA where a contractor talks into their phone and AI routes everything to the right place: materials to shopping lists grouped by store, dates to Google Calendar color-coded by job, tasks, or per-job notes.

Fuzzy entity resolution matches phrases like "the Kellogg job" or "Chris's place" to the correct project by injecting the active job list into the system prompt. A Claude Vision path turns site photos plus voice into structured scope with materials, code flags, and cost estimates. Meeting recorder runs Deepgram diarized transcription into summarized notes. Voice input queues offline for dead zones.

Crew management with row-level security so employees see only assigned jobs. Subcontractor directory with per-job placement.

Free AI resume diagnostic. Paste a resume and job description, get a match score (0 to 100) across four dimensions (keyword match, experience relevance, trajectory fit, ATS parsing), a GO / FIX FIRST / PASS verdict, and a downloadable ATS-safe Word doc rewritten to mirror the job listing.

The core constraint: no fabrication. A three-layer guard enforces it. System prompts restrict the model to user-provided evidence. A five-question intake surfaces real stories before tailoring runs. A separate fact-check pass verifies every rewritten bullet against the original resume, regenerating failures with per-claim feedback.

Company-targeting mode: paste a URL instead of a job description and get a tailored resume with positioning angle. Cost-tiered model use (Haiku for diagnosis, Sonnet for generation). Word-level diff shows what changed.

Chrome extension that captures open tabs, auto-categorizes them, and lets users annotate, snooze, and snapshot them into restorable sessions. Built for knowledge workers who keep dozens of tabs open as external memory and lose time re-orienting after each context switch. Where competitors save URLs, TabSquirrel saves context: what the user was doing and what comes next.

The build pairs a Manifest V3 extension with a React SPA over a postMessage bridge and externally_connectable, avoiding a hardcoded extension ID. Free-tier limits are enforced at the database layer through a Postgres BEFORE INSERT trigger using SECURITY DEFINER. A split tabs/tab_meta model with URL normalization handles dedup on import. Every table runs row-level security keyed to auth.uid().

Automated pipeline that ingests sales call transcripts, extracts structured coaching insights, grades each call A through F, and verifies every product claim against a 672-document knowledge base embedded in Supabase pgvector.

Built for the sales director at Tracker Products. Two AI stages: Gemini 2.0 Flash returns strict JSON (agency, persona, objections, competitor mentions, rubric score); a LangChain compliance agent checks each claim against the knowledge base, returning PASS, FLAG, or REVIEW with evidence quotes.

Reps contest results via tracking codes in Slack, where a router resolves disputes. Reliability includes scheduled batch runs, two-layer deduplication, malformed-JSON repair, and two-step archiving.

Internal AI knowledge base for Tracker Products. The system turns SOPs, product documentation, and institutional knowledge into a searchable operating layer for the team.

Three modules: War Room for RAG chat, Knowledge Factory for document ingestion, and Ops Manual for SOP browsing. It was built for a real company with real team usage, not as a demo.

The key architecture decision was consolidating auth, row-level security, storage, and vector search inside Supabase instead of splitting the system across separate tools.

Persistent memory system for AI work. Open Brain captures decisions, observations, people notes, references, and tasks from AI sessions so future agents can search and reuse context instead of starting over.

The system separates tasks from thoughts, deduplicates pending work by semantic similarity, supports typed relationships between thoughts, and exposes read and write tools through an MCP server.

This is the infrastructure I use daily to keep AI work connected across sessions, tools, and projects.

Publishing system for a weekly AI practitioner roundtable. It turns call transcripts and edited recaps into a static web archive that updates through the repo and deploys on Vercel.

The workflow pairs transcript files with recap markdown by date, rebuilds the public archive, and keeps the publishing path repeatable without hand-editing pages.

It is a practical example of content operations as infrastructure: capture the source, turn it into a useful artifact, publish it, and keep the process maintainable.

Credential management workflow that bridges 1Password secrets into project-specific development environments without copying keys into code, chats, or local notes.

The system uses reusable environment templates, shell helpers, biometric 1Password CLI authentication, and project setup patterns so AI-heavy development work can access the right credentials safely.

It is not a flashy app. It is the operating layer that keeps multi-project AI work from becoming a security mess.

Interactive tool that audits recurring business processes for hidden cost. Enter a process with its frequency, headcount, duration, and what mistakes it prevents. The app returns a Keep, Compress, or Cut verdict with rationale and a specific recommended action.

Runs as a deterministic rules engine with no AI calls and no network requests. A cost model converts inputs to monthly team-hours using per-frequency multipliers, and a 40-hour threshold separates Compress from Keep. Rationale and actions generate from live numbers rather than canned text. Processes accumulate into a ranked summary table with clipboard export. Everything runs client-side, nothing saved or transmitted.

Built for operators and team leads who accumulate standing meetings, reviews, and reports that nobody re-evaluates because the cost sits in payroll hours and stays invisible.