The FDE Portfolio: 5 High-Velocity Prototypes That Prove You Ship in Chaos

The Portfolio Paradox: Why CRUD Apps Don't Cut It

A standard software engineering portfolio proves you can build. A Forward Deployed Engineer (FDE) portfolio proves you can build while the building is on fire.

The market is bifurcating. Standard SWE loops ask for LeetCode and system design. The FDE interview loop—which unlocks total compensation packages often reaching $250K–$450K+ at top AI labs—asks a different question: "Show me a time you walked into a messy, legacy enterprise and shipped value in 48 hours."

If your portfolio is a beautifully architected To-Do app with 100% test coverage, you are signaling safety. FDEs signal velocity. We don't just tolerate chaos; we find signal in the noise. The following five prototypes are designed to fill your portfolio with evidence of high-agency, high-velocity shipping.

For a deeper dive into what hiring managers actually look for in these scenarios, review The FDE Interview Loop: Concrete Prep Scenarios for Builders Who Ship.

Prototype 1: The CRM Ghostbuster (Real-Time Data Unification)

The Scenario: A sales team lives in Salesforce, but critical product usage data lives in a separate PostgreSQL database. Deals are lost because sales reps don't know if a prospect is a power user or a ghost.

The FDE Move: Don't wait for a data engineering pipeline. Build a lightweight Chrome Extension that detects the current Salesforce Contact ID and queries the product DB via a thin FastAPI wrapper.

The Signal: This demonstrates you don't need a perfect API. You'll screen-scrape the DOM for the Salesforce ID if you have to. You understand that 80% of FDE work is gluing systems that were never meant to talk.

Stack: Chrome Extension Manifest V3, FastAPI, SQLAlchemy, and a single docker-compose.yml.

Key Metric to Showcase: "Reduced 'swivel chair' data lookups from 45 seconds to 0ms, directly embedded in the workflow."

Prototype 2: The Legacy Log Whisperer (Regex-to-API Bridge)

The Scenario: A massive logistics company has 15 years of critical operational logs stored in unstructured .txt files on a Windows Server 2008 machine. No API exists.

The FDE Move: Write a Python watchdog script that tails these logs, applies regex to extract structured events (e.g., shipment delays), and pushes them to a Slack webhook or a modern observability tool.

The Signal: This proves you don't need structured JSON to provide value. You are willing to meet the customer where they are—even if "where they are" is a deprecated SMB share. This is the essence of the FDE Product & Engineering Post-Sale Collaboration.

Code Block:

import re
from watchdog.observers import Observer
from watchdog.events import FileSystemEventHandler

PATTERN = r"(ERROR|CRITICAL): (.*)"

class LogHandler(FileSystemEventHandler):
    def on_modified(self, event):
        with open(event.src_path, 'r') as f:
            lines = f.readlines()
            for line in lines[-10:]: # tail the last 10 lines
                match = re.search(PATTERN, line)
                if match:
                    send_slack_alert(match.group(2))

Prototype 3: The Air-Gapped Inference Engine (Local RAG on Sensitive Docs)

The Scenario: A defense contractor or legal firm wants to query their internal policy documents, but data cannot leave the premises. Cloud APIs (OpenAI, Anthropic) are blocked.

The FDE Move: Deploy a local Retrieval-Augmented Generation (RAG) pipeline using Ollama (Mistral or Llama 3) and ChromaDB. Index thousands of PDFs locally. Build a Streamlit UI that runs entirely on an air-gapped machine.

The Signal: This isn't a demo; it's an operational constraint. It shows you understand the hardware and trust boundaries required for enterprise AI, similar to the constraints explored in Why LLMs Remain Useless Against AES: The Cryptographic Wall.

Stack: Ollama, LangChain, ChromaDB, PyMuPDF, Streamlit.

Prototype 4: The "Swivel Chair" Killer (Browser Automation for Ops)

The Scenario: An operations team manually re-enters data from a modern dashboard into a legacy AS/400 terminal (green screen) for compliance reasons. This takes 4 hours daily.

The FDE Move: Use Playwright or Puppeteer to automate the legacy terminal emulator (which is often just a web-based Java applet). Script the copy-paste flow between the modern Grafana dashboard and the legacy system.

The Signal: You don't scoff at legacy tech; you automate it. This is pure cost-savings engineering. It shows you understand that "digital transformation" rarely means ripping out the mainframe; it means bridging it.

Tool Note: Use pyautogui as a fallback if the terminal is a native app rather than browser-based. FDEs play dirty (in a good way).

Prototype 5: The Multi-Agent Procurement Auditor (Planning, Search, Write)

The Scenario: A procurement team spends weeks manually researching vendors, checking compliance, and writing justification memos.

The FDE Move: Build a multi-agent system where one agent plans the research strategy, a second searches the web (Tavily) and internal PDFs, and a third synthesizes a compliance brief.

This is a direct application of the patterns found in Build a Multi-Agent Research Assistant That Plans, Searches, and Writes a Brief with Gemini.

The Signal: This shows you can orchestrate non-deterministic AI workflows. It proves you understand the "builder-seller" archetype: you aren't just hooking up an API; you are compressing a human workflow from 40 hours to 40 seconds.

Architecture Deep-Dive: The Ghostbuster Flow

When presenting Prototype 1 (The CRM Ghostbuster), a text description isn't enough. You need to show the data flow. Instead of a static image, use a programmatic diagram. Here is the flow for the real-time data unification Chrome Extension:

How to Present Chaos to a Hiring Manager

Don't just list these on GitHub. Write a "Field Report" for each.

1. The Blockade: What was the messy reality? (e.g., "No API key was available; the IT team was on vacation.")
2. The Breach: The ugly, practical code you wrote. Ship the script, not the framework.
3. The Value: Time saved, revenue influenced, or decisions enabled.

FDE Coach often sees engineers fail interviews not because they lack skill, but because they sanitize their stories. Keep the duct tape visible. A script that parses a PDF by converting it to text and using split() is often a better FDE signal than a perfectly abstracted class hierarchy.

If you need to upskill in the specific AI patterns used in Prototype 5, moving beyond simple chatbots to autonomous agents is the single highest-leverage skill you can acquire right now.

FAQ: Portfolio Logistics

Q: Do I need a polished frontend for these prototypes? A: No. A Gradio or Streamlit interface is perfectly acceptable. If you spend 8 hours on CSS, you are missing the point. The value is the logic that bridges the gap between the legacy system and the modern world.

Q: What if I don't have access to a "legacy" system? A: Simulate it. Download a 10-year-old open-source ERP system (like Odoo Community v8) and treat it as the legacy monolith. Create a CSV export that looks like a mainframe dump. The constraint is the point.

Q: How many prototypes should I show in an interview? A: Have 5 ready, but deep-dive on 2. Pick the two that best match the customer profile of the company you are interviewing with.