Proof of Concept — Technical Plan
Client: Taleb Alashkar
Prepared by: Faizan Ali | June 2026
A screening tool that checks Texas attorney advertising (blog posts, websites, social media, directory listings) for compliance with the Texas Disciplinary Rules of Professional Conduct (Rules 7.01–7.05). The user pastes a URL, the system fetches the content, sends it to an LLM with the full disciplinary rules context, and returns a GREEN / YELLOW / RED verdict with specific findings, rule citations, and suggested fixes.
Important framing: This is an advisory screening tool, not a compliance certifier. Every output includes a disclaimer that it does not constitute legal advice and the burden of compliance remains on the attorney.
POC scope: Local Windows app for live screen-share demos. No cloud hosting, no auth. Taleb pastes a link, gets results.
Paste URL → Fetch & extract page content → Send to LLM with Texas Rules 7.01–7.05 → LLM returns structured findings (violations, severity, rule citations, confidence, suggested rewrites) → Display color-coded verdict with per-finding breakdown
The LLM does all the analysis — both bright-line violations (missing disclaimers, prohibited terms) and contextual/fuzzy issues (misleading implications, unjustified expectations). No separate rule engine or regex system needed. The disciplinary rules are provided as context in the system prompt, and the LLM identifies violations, cites the specific rule, and suggests compliant alternatives.
| Category | Examples | Severity |
|---|---|---|
| Bright-line / Structural | Missing "ADVERTISEMENT" marking (7.03), no responsible attorney name (7.04), "specialist" without TBLS cert (7.02), superlative trade names (7.01(c)) | RED |
| Pattern-risk | Guarantees ("we always win"), unsubstantiated stats, testimonials without disclaimers (7.01 + Cmt. 10), bilingual disclaimer mismatch (7.02) | YELLOW → RED |
| Contextual / Fuzzy | "Materially misleading as a whole," "unjustified expectations," omission of necessary facts (7.01(a)) | YELLOW only |
| Source | How |
|---|---|
| Blog posts | Paste URL → auto-fetch |
| Law firm websites | Paste URL → auto-fetch |
| Social media posts | Paste URL or paste text directly |
| Online directories (Avvo, FindLaw) | Paste URL → auto-fetch |
| Google Ads / PPC copy | Paste text directly |
| Component | Technology | Purpose |
|---|---|---|
| Language | Python 3.12+ | Primary backend |
| Web Framework | Django 6+ | Application framework + REST API |
| Database | SQLite | Scan history — zero config |
| Validation | Pydantic 2.x | LLM output validation |
| URL Fetching | httpx + BeautifulSoup | Fetch and extract page content from any URL |
Two providers configured. If one runs out of tokens or hits rate limits, switch to the other.
| Option | Provider | How It Works | Cost |
|---|---|---|---|
| A (Default) | OpenAI API | Direct API calls to GPT-4.1. Structured JSON output via response_format. | Pay-as-you-go (~$0.02-0.05/analysis) |
| B (Fallback) | Codex CLI | Runs codex exec as a subprocess with the analysis prompt. Uses existing Codex subscription. | Free tier available. Plus ($20/mo), Pro (from $100/mo). |
| Component | Technology | Purpose |
|---|---|---|
| Framework | Next.js 16 | Web application (served locally) |
| UI | shadcn/ui + Tailwind CSS | Polished, demo-ready interface |
| Component | Purpose |
|---|---|
| Windows machine | Taleb's local machine |
start.bat | One-click: launches Django + Next.js together |
| Item | Cost/mo | Notes |
|---|---|---|
| AI API (OpenAI or Codex) | $5-15 | Demo volume only. ~$0.02-0.05 per URL analyzed. |
| Everything else | $0 | All local, all free |
| TOTAL | $5-15/month |
| # | Feature | Hours | Key Details |
|---|---|---|---|
| 1 | URL Fetching & Content Extraction | 2-3 | Fetch any URL (blog, website, directory), extract clean text content via BeautifulSoup. Handle common edge cases (JS-rendered pages, paywalls) gracefully with error messages. |
| 2 | AI Analysis & Prompt Engineering | 4-6 | System prompt with full Texas Rules 7.01–7.05. Structured JSON output (findings, severity, confidence, rule citations, suggested rewrites, overall verdict). Dual-provider support (Claude API + Codex CLI). Pydantic validation of AI response. |
| 3 | Frontend UI | 8-10 | Input form (URL/text toggle), loading state, color-coded verdict banner (GREEN/YELLOW/RED), per-finding cards with rule citations and severity badges, suggested rewrite section, disclaimer footer. Polished for demo. |
| 4 | Scan History | 1-2 | List of past scans with verdict badges. Click to view full results. Shows well in demo. |
| 5 | Setup & Config | 1-2 | start.bat, .env for API keys, setup instructions |
| Subtotal | 16-23 | ||
| Testing + Bug Fixes | 2-4 | ||
| Project Management | 1-2 | ||
| PHASE 1 TOTAL | 19-29 |
* Hours estimated with AI-assisted development (Claude Code).
| Metric | Value |
|---|---|
| Phase 1 Hours (POC) | 19 - 29 hours |
| Hourly Rate | $38/hour |
| Phase 1 Cost (POC) | $722 - $1,102 |
| Timeline | ~1 week |
| Monthly Running Cost | $5 - $15 (AI API only) |
Prepared by Faizan Ali