Projects

🧠 AI Automation & Data Engineering Projects

A collection of builds spanning two disciplines: AI-powered automation — LLM-driven document processing, agentic workflows, and human-in-the-loop approval systems — and end-to-end data engineering, carrying raw data from source through BigQuery and dbt to live production dashboards in Power BI.

📊 Medicare Data Pipeline: Raw Data to Production Dashboard

Type: Data Engineering / Cloud Data Pipeline Stack: Google Cloud Storage, BigQuery, dbt, Power BI Code: View on GitHub

A real, working pipeline built on public CMS Medicare provider/payment data — de-identified, public-domain data, no PII — carried end to end from raw CSV to a live production dashboard. Built specifically to demonstrate cloud-native data warehousing and transformation (BigQuery + dbt), not just a script that reads a CSV.

Highlights

Raw CMS provider/service data (millions of rows) landed in Cloud Storage, loaded to BigQuery, modeled with dbt
Staging layer normalizes and types raw fields; production marts aggregate to provider- and procedure-level summaries
dbt-defined tests (not_null, unique) enforce data quality on every run
Power BI connects directly to the BigQuery production marts — no manual export/import step

🧩 Workflow Diagram

📊 Medicare Data Pipeline — Diagram

%%{init: {'flowchart': { 'htmlLabels': true, 'wrap': true, 'nodeSpacing': 60, 'rankSpacing': 80 }}}%%
flowchart LR
  subgraph Ingest [📥 Ingest]
    A1["(1) data.cms.gov<br/>raw CSV download"]
    A2["(2) Cloud Storage<br/>raw landing zone"]
    A1 --> A2
  end

  A2 --> B1["(3) BigQuery<br/>bq load raw table"]

  subgraph Transform [🧱 dbt Transform]
    B1 --> C1["(4) Staging model<br/>clean, rename, type-cast"]
    C1 --> D1["(5) Mart: provider summary<br/>aggregated by NPI"]
    C1 --> D2["(6) Mart: procedure summary<br/>aggregated by HCPCS + state"]
    D1 --> E1["(7) dbt test<br/>not_null / unique checks"]
    D2 --> E1
  end

  subgraph Serve [📈 Serve]
    E1 --> F1["(8) Power BI<br/>connects to BigQuery marts"]
    F1 --> F2["(9) Publish to web<br/>live embedded dashboard"]
  end

Workflow Steps

Source: Download the CMS “Medicare Physician & Other Practitioners - by Provider and Service” CSV from data.cms.gov
Land: Upload the raw file to a Cloud Storage bucket — the raw, untouched landing zone
Load: bq load the CSV into a raw BigQuery table, no transformation yet
Stage: dbt staging model cleans column names, casts types, filters obvious nulls — no business logic
Model (provider): dbt mart aggregates to one row per provider (NPI) — total services, beneficiaries, Medicare payments
Model (procedure): dbt mart aggregates to one row per (procedure code, state) — cost and volume by geography
Test: dbt’s built-in test framework enforces not_null/unique constraints on key columns before the marts are trusted downstream
Connect: Power BI connects directly to the BigQuery production marts — no CSV export step in between
Publish: Power BI’s “Publish to web” generates a live, filterable embed for the portfolio site

🎬 Try it live

This is the actual production Power BI dashboard, built on the pipeline above — filter and explore it directly.

📈 Superstore Sales & Profitability Dashboard

Type: Data Engineering / Cloud Data Pipeline Stack: BigQuery, dbt, Power BI Code: View on GitHub

A complete, working pipeline built on the Sample Superstore dataset — sales, profit, and discount data across product categories, regions, and customer segments — carried from a raw CSV straight through to a live production dashboard. Deliberately chosen after an earlier attempt (GA4 e-commerce data) turned out to have deliberately-stripped pricing fields — this dataset is complete by design, letting the analysis focus on the actual business question instead of working around missing data.

Highlights

Direct CSV upload straight into BigQuery — no Cloud Storage step needed at this scale
dbt staging layer normalizes raw column names (including literal spaces/hyphens from the source CSV) before any business logic runs
Two production marts: category/region profitability, and segment/state performance
Reveals a real, visible pattern in the data — heavier discounting correlates with lower profit margin, and it hits some categories far harder than others
Grouping logic applied defensively from the start (using ANY_VALUE() for descriptive fields rather than a fragile composite key) — a lesson carried over from an earlier project rather than learned the hard way twice

🧩 Workflow Diagram

📈 Superstore Pipeline — Diagram

%%{init: {'flowchart': { 'htmlLabels': true, 'wrap': true, 'nodeSpacing': 60, 'rankSpacing': 80 }}}%%
flowchart LR
  subgraph Ingest [📥 Ingest]
    A1["(1) Sample Superstore CSV<br/>direct download"]
    A2["(2) BigQuery<br/>direct upload, no GCS needed"]
    A1 --> A2
  end

  subgraph Transform [🧱 dbt Transform]
    A2 --> B1["(3) Staging model<br/>clean, rename, type-cast"]
    B1 --> C1["(4) Mart: category/region<br/>sales, profit, margin, discount"]
    B1 --> C2["(5) Mart: segment/state<br/>profit by segment + geography"]
    C1 --> D1["(6) dbt test<br/>not_null / unique checks"]
    C2 --> D1
  end

  subgraph Serve [📈 Serve]
    D1 --> E1["(7) Power BI<br/>connects to BigQuery marts"]
    E1 --> E2["(8) Publish to web<br/>live embedded dashboard"]
  end

Workflow Steps

Source: Download the Sample Superstore CSV (complete, no missing-by-design fields)
Load: Direct upload into BigQuery — small enough to skip Cloud Storage entirely
Stage: dbt staging model cleans column names (including literal spaces/hyphens from the raw headers) and casts types
Model (category/region): dbt mart aggregates sales, profit, quantity, discount, and margin by category, sub-category, and region
Model (segment/state): dbt mart aggregates the same metrics by customer segment and state
Test: dbt’s test framework enforces not_null/unique constraints before the marts are trusted downstream
Connect: Power BI connects directly to the BigQuery production marts
Publish: Power BI’s “Publish to web” generates a live, filterable embed for the portfolio site

🎬 Try it live

The actual production Power BI dashboard, built on the pipeline above — filter by segment and explore it directly.

🗳️ 2026 Congressional Campaign Finance Dashboard

Type: Data Engineering / Cloud Data Pipeline Stack: BigQuery, dbt, Power BI Code: View on GitHub

A pipeline built on genuinely live, current data — real 2025-2026 federal election cycle campaign finance filings, downloaded directly from the FEC’s official bulk data files and carried through to a live production dashboard. Unlike the other pipelines in this portfolio, this one tracks an active, ongoing dataset that updates as new filings are received, rather than a static historical snapshot.

Highlights

Built against the FEC’s exact real column schema, verified directly against the source file before writing any SQL — no guess-and-fix iteration needed
Reveals a genuinely contrasting story: House races raise more total dollars, but Senate candidates raise more per-candidate on average — visible side by side in two deliberately paired charts
Surfaces the well-documented incumbency advantage clearly: incumbents outraise both open-seat and challenger candidates by a wide margin
Party codes normalized at the source layer (e.g., Minnesota’s “DFL” folded into “DEM”) so downstream charts tell an accurate national story

🧩 Workflow Diagram

🗳️ FEC Campaign Finance Pipeline — Diagram

%%{init: {'flowchart': { 'htmlLabels': true, 'wrap': true, 'nodeSpacing': 60, 'rankSpacing': 80 }}}%%
flowchart LR
  subgraph Ingest [📥 Ingest]
    A1["(1) FEC.gov<br/>candidate_summary_2026.csv"]
    A2["(2) BigQuery<br/>direct upload, no GCS needed"]
    A1 --> A2
  end

  subgraph Transform [🧱 dbt Transform]
    A2 --> B1["(3) Staging model<br/>clean, rename, normalize party codes"]
    B1 --> C1["(4) Mart: fundraising by party/office<br/>totals + per-candidate averages"]
    B1 --> C2["(5) Mart: candidate leaderboard<br/>one row per candidate"]
    C1 --> D1["(6) dbt test<br/>not_null / unique checks"]
    C2 --> D1
  end

  subgraph Serve [📈 Serve]
    D1 --> E1["(7) Power BI<br/>connects to BigQuery marts"]
    E1 --> E2["(8) Publish to web<br/>live embedded dashboard"]
  end

Workflow Steps

Source: Download the FEC’s official “candidate_summary_2026.csv” bulk data file for the 2025-2026 election cycle
Load: Direct upload into BigQuery — small enough to skip Cloud Storage entirely
Stage: dbt staging model cleans column names, casts types, and normalizes party affiliation codes (e.g., DFL → DEM) and office codes (H/S/P → House/Senate/President)
Model (fundraising): dbt mart aggregates receipts, disbursements, and cash-on-hand by party and office, including both totals and per-candidate averages
Model (leaderboard): dbt mart produces one row per candidate, grouped defensively by the reliable candidate_id field
Test: dbt’s test framework enforces not_null/unique constraints before the marts are trusted downstream
Connect: Power BI connects directly to the BigQuery production marts
Publish: Power BI’s “Publish to web” generates a live, filterable embed for the portfolio site

🎬 Try it live

The actual production Power BI dashboard, built on genuinely live 2026 election cycle data — explore it directly.

🏀 NBA Team Performance & Shooting Trends

Type: Data Engineering / Cloud Data Pipeline Stack: BigQuery, dbt, Power BI Code: View on GitHub

A pipeline built on real NBA team game data spanning 2010-2024 — fourteen regular seasons carried from a public, actively-maintained GitHub dataset through to a live production dashboard, tracking the sport’s most significant modern tactical shift: the rise of the three-point shot.

Highlights

Built against a fully-documented public schema, verified directly against the source repo before writing any SQL
League-wide shooting trends reveal the well-documented “3-point revolution” — average 3-point attempts per team climbing steadily across the full 2010-2024 span
Team-level win percentage, scoring, and point differential tracked season over season, filterable by team
PLUS_MINUS used directly for point differential, avoiding an unnecessary self-join against opponent data

🧩 Workflow Diagram

🏀 NBA Team Stats Pipeline — Diagram

%%{init: {'flowchart': { 'htmlLabels': true, 'wrap': true, 'nodeSpacing': 60, 'rankSpacing': 80 }}}%%
flowchart LR
  subgraph Ingest [📥 Ingest]
    A1["(1) GitHub<br/>regular_season_totals_2010_2024.csv"]
    A2["(2) BigQuery<br/>direct upload, no GCS needed"]
    A1 --> A2
  end

  subgraph Transform [🧱 dbt Transform]
    A2 --> B1["(3) Staging model<br/>clean, rename, type-cast"]
    B1 --> C1["(4) Mart: team season summary<br/>win %, scoring, shooting efficiency"]
    B1 --> C2["(5) Mart: league shooting trends<br/>3-point attempts by season"]
    C1 --> D1["(6) dbt test<br/>not_null / unique checks"]
    C2 --> D1
  end

  subgraph Serve [📈 Serve]
    D1 --> E1["(7) Power BI<br/>connects to BigQuery marts"]
    E1 --> E2["(8) Publish to web<br/>live embedded dashboard"]
  end

Workflow Steps

Source: Download “regular_season_totals_2010_2024.csv” from the publicly maintained NocturneBear/NBA-Data-2010-2024 GitHub repository
Load: Direct upload into BigQuery — no Cloud Storage step needed at this file size
Stage: dbt staging model cleans column names and casts types against the source’s own documented schema
Model (team season): dbt mart aggregates wins, losses, win percentage, scoring, and shooting efficiency by team and season
Model (league trends): dbt mart aggregates league-wide shooting and scoring averages by season, surfacing the 3-point attempt trend
Test: dbt’s test framework enforces not_null/unique constraints before the marts are trusted downstream
Connect: Power BI connects directly to the BigQuery production marts
Publish: Power BI’s “Publish to web” generates a live, filterable embed for the portfolio site

🎬 Try it live

The actual production Power BI dashboard, built on the pipeline above — filter by team and explore fourteen seasons of shooting trends directly.

🎯 March Madness ML: 2025 NCAA Bracket Predictions

Type: Machine Learning / Predictive Modeling Stack: Python, pandas, scikit-learn, XGBoost Code: View on GitHub

A machine learning pipeline trained blind on NCAA tournament data through 2024 — from the official Kaggle “March Machine Learning Mania” competition dataset, frozen right before Selection Sunday 2025 — then scored against the real 2025 tournament results once they’d actually happened. Both Men’s and Women’s brackets.

Highlights

Feature engineering across 7,981 (Men’s) and 5,602 (Women’s) team-seasons of real box score data
Two models compared head-to-head (Logistic Regression, XGBoost) with proper time-based validation — trained on older seasons, tested on held-out recent ones, not randomly shuffled
Real-world validation: predicted the entire 2025 bracket before it played out, then checked against actual results — 12/14 (85.7%) correct across the Elite Eight through Championship, including a perfect 7/7 on the Women’s side

🧩 Workflow Diagram

🎯 March Madness ML — Diagram

%%{init: {'flowchart': { 'htmlLabels': true, 'wrap': true, 'nodeSpacing': 60, 'rankSpacing': 80 }}}%%
flowchart LR
  subgraph Features [📥 Feature Engineering]
    A1["(1) Kaggle dataset<br/>box scores through 2024"]
    A2["(2) Per-team-season stats<br/>win%, shooting, rebounds, etc."]
    A1 --> A2
  end

  subgraph Training [🧠 Model Training]
    A2 --> B1["(3) Matchup training data<br/>symmetric augmentation"]
    B1 --> B2["(4) Train + validate<br/>LogReg vs XGBoost, time-split"]
  end

  subgraph Predict [🔮 2025 Predictions]
    B2 --> C1["(5) Every possible matchup<br/>among 68 tournament teams"]
    C1 --> C2["(6) Simulate real bracket<br/>round by round"]
  end

  subgraph Validate [✅ Real-World Scoring]
    C2 --> D1["(7) Fetch actual 2025 results<br/>after the tournament happened"]
    D1 --> D2["(8) Score predictions<br/>vs. real outcomes"]
  end

Workflow Steps

Source: Kaggle’s official “March Machine Learning Mania 2025” competition dataset — frozen before the 2025 tournament, so 2025 results genuinely weren’t in the training data
Feature engineering: aggregate each team’s season stats (win %, scoring, shooting efficiency, rebounding, turnovers) from game-level box scores
Training data: historical tournament games become matchup rows, symmetrically augmented (each game contributes one “Team A wins” and one “Team B wins” row) to avoid team-order bias
Train & validate: Logistic Regression and XGBoost, evaluated on held-out recent tournament seasons — not a random split, which would leak future information
Generate 2025 predictions: win probability for every possible pairing among the 68 tournament teams
Simulate the bracket: walk the real slot structure (First Four through Championship) round by round, advancing the higher-probability team at each step
Fetch real results: once the 2025 tournament had actually happened, pull the real outcomes from published sources
Score: compare blind predictions against real results — 12/14 correct across the Elite Eight through Championship

🏀 Predictions vs. What Actually Happened

Trained blind on data through 2024, then scored against the real 2025 tournament outcomes.

Backtested Accuracy (Men's)

67.9%

vs. 68.3% seed-only baseline

Backtested Accuracy (Women's)

81.3%

vs. 77.2% seed-only baseline

2025 Real Results (Combined)

12 / 14

Elite Eight through Championship

🏀 Men's Bracket — 5/7 correct

Round	Our Pick	Actual	✓
Elite 8	Florida	Florida	✓
Elite 8	Duke	Duke	✓
Elite 8	Auburn	Auburn	✓
Elite 8	Houston	Houston	✓
Final Four	Houston	Houston	✓
Final Four	Auburn	Florida	✗
Champion	Houston	Florida	✗

Correctly called all four Elite Eight winners, matching the historic all-#1-seed Final Four. Picked the actual national runner-up as champion.

🏀 Women's Bracket — 7/7 correct

Round	Our Pick	Actual	✓
Elite 8	UConn	UConn	✓
Elite 8	S. Carolina	S. Carolina	✓
Elite 8	UCLA	UCLA	✓
Elite 8	Texas	Texas	✓
Final Four	S. Carolina	S. Carolina	✓
Final Four	UConn	UConn	✓
Champion	UConn	UConn	✓

Perfect: all four Elite Eight winners, both Final Four semifinal winners, and the eventual champion — all called correctly, blind, before the tournament played out.

Scoring covers the Elite Eight through Championship (the rounds that decide the tournament), sourced from published game results. Earlier rounds (First/Second Round, Sweet 16) weren't individually re-verified against game-by-game results for this summary.

🎯 Full Predicted Bracket — 2025

Every round, every predicted winner. Green/red borders mark picks verified against real results (Elite Eight through Championship); gray borders are earlier-round predictions not individually re-verified against game-by-game results.

🏀 Men's Bracket

First Four

American Univ

Mt St Mary's

American Univ

Texas

Xavier

Texas

North Carolina

San Diego St

Alabama St

St Francis PA

Alabama St

Round of 64

Duke

American Univ

Duke

Alabama

Robert Morris

Alabama

Wisconsin

Montana

Wisconsin

Arizona

Akron

Arizona

Oregon

Liberty

Oregon

BYU

VCU

St Mary's CA

Vanderbilt

St Mary's CA

Mississippi St

Baylor

Mississippi St

Houston

SIUE

Houston

Tennessee

Wofford

Tennessee

Kentucky

Troy

Kentucky

Purdue

High Point

Purdue

Clemson

McNeese St

Clemson

Illinois

Texas

Illinois

UCLA

Utah St

UCLA

Gonzaga

Georgia

Gonzaga

Auburn

Alabama St

Auburn

Michigan St

Bryant

Michigan St

Iowa St

Lipscomb

Iowa St

Texas A&M

Yale

Texas A&M

Michigan

UC San Diego

Mississippi

San Diego St

Mississippi

Marquette

New Mexico

Marquette

Louisville

Creighton

Louisville

Florida

Norfolk St

Florida

St John's

NE Omaha

St John's

Texas Tech

UNC Wilmington

Texas Tech

Maryland

Grand Canyon

Maryland

Memphis

Colorado St

Memphis

Missouri

Drake

Missouri

Kansas

Arkansas

Kansas

Connecticut

Oklahoma

Connecticut

Round of 32

Duke

Mississippi St

Duke

Alabama

St Mary's CA

Alabama

Wisconsin

VCU

Wisconsin

Arizona

Oregon

Arizona

Houston

Gonzaga

Houston

Tennessee

UCLA

Tennessee

Kentucky

Illinois

Kentucky

Purdue

Clemson

Auburn

Louisville

Auburn

Michigan St

Marquette

Michigan St

Iowa St

Mississippi

Iowa St

Texas A&M

UC San Diego

Texas A&M

Florida

Connecticut

Florida

St John's

Kansas

St John's

Texas Tech

Missouri

Texas Tech

Maryland

Memphis

Maryland

Sweet 16

Duke

Arizona

Duke

Alabama

Wisconsin

Houston

Clemson

Houston

Tennessee

Kentucky

Tennessee

Auburn

Texas A&M

Auburn

Michigan St

Iowa St

Florida

Maryland

Florida

St John's

Texas Tech

St John's

Elite Eight

Duke

Wisconsin

Duke ✓

Houston

Tennessee

Houston ✓

Auburn

Iowa St

Auburn ✓

Florida

St John's

Florida ✓

Final Four

Duke

Houston

Houston ✓

Auburn

Florida

Auburn ✗

Championship

Houston

Auburn

Houston ✗

🏀 Women's Bracket

First Four

Columbia

Washington

Columbia

Iowa St

Princeton

Iowa St

High Point

William & Mary

High Point

Southern Univ

UC San Diego

Round of 64

South Carolina

Tennessee Tech

South Carolina

Duke

Lehigh

Duke

North Carolina

Oregon St

North Carolina

Maryland

Norfolk St

Maryland

Alabama

WI Green Bay

Alabama

West Virginia

Columbia

West Virginia

Vanderbilt

Oregon

Vanderbilt

Utah

Indiana

Utah

Texas

High Point

Texas

TCU

F Dickinson

TCU

Notre Dame

SF Austin

Notre Dame

Ohio St

Montana St

Ohio St

Tennessee

South Florida

Tennessee

Michigan

Iowa St

Michigan

Louisville

Nebraska

Louisville

Illinois

Creighton

UCLA

UC San Diego

UCLA

NC State

Vermont

NC State

LSU

San Diego St

LSU

Baylor

Grand Canyon

Baylor

Mississippi

Ball St

Mississippi

Florida St

George Mason

Florida St

Michigan St

Harvard

Michigan St

Richmond

Georgia Tech

Richmond

USC

UNC Greensboro

USC

Connecticut

Arkansas St

Connecticut

Oklahoma

FGCU

Oklahoma

Kentucky

Liberty

Kentucky

Kansas St

Fairfield

Kansas St

Iowa

Murray St

Iowa

Oklahoma St

S Dakota St

Oklahoma St

California

Mississippi St

Round of 32

South Carolina

Utah

South Carolina

Duke

Vanderbilt

Duke

North Carolina

West Virginia

North Carolina

Maryland

Alabama

Texas

Creighton

Texas

TCU

Louisville

TCU

Notre Dame

Michigan

Notre Dame

Ohio St

Tennessee

Ohio St

UCLA

Richmond

UCLA

NC State

Michigan St

NC State

LSU

Florida St

LSU

Baylor

Mississippi

Baylor

USC

Mississippi St

USC

Connecticut

Oklahoma St

Connecticut

Oklahoma

Iowa

Oklahoma

Kentucky

Kansas St

Sweet 16

South Carolina

Alabama

South Carolina

Duke

North Carolina

Duke

Texas

Ohio St

Texas

TCU

Notre Dame

TCU

UCLA

Baylor

UCLA

NC State

LSU

USC

Kansas St

USC

Connecticut

Oklahoma

Connecticut

Elite Eight

South Carolina

Duke

South Carolina ✓

Texas

TCU

Texas ✓

UCLA

LSU

UCLA ✓

USC

Connecticut

Connecticut ✓

Final Four

South Carolina

Texas

South Carolina ✓

UCLA

Connecticut

Connecticut ✓

Championship

South Carolina

Connecticut

Connecticut ✓

⚙️ Contract Processing Bot

Type: Document Intelligence
Stack: Zapier, OpenAI, Google Drive, Gmail, Slack, Google Sheets

This bot automates contract intake and review—extracting key clauses, identifying risks, logging results, and routing approvals.

Highlights

80–90% reduction in contract triage time
Real-time Slack alerts for risky clauses
Seamless Google Drive + Sheets tracking

🧩 Workflow Diagram

⚙️ Contract Processing Bot — Diagram

%%{init: {'flowchart': { 'htmlLabels': true, 'wrap': true, 'nodeSpacing': 60, 'rankSpacing': 80 }}}%%
flowchart LR
  subgraph Intake [📥 Intake]
    A1["(1) Gmail Trigger<br/>new attachment 'contract'"]
    A2["(2) Slack<br/>intake notification"]
    A3{"(3) Filter<br/>contract file?"}
    A1 --> A2 --> A3
  end

  A3 -- Yes --> A4["(4) Save to Drive<br/>/Contracts/Incoming"]
  A3 -- No  --> R1["Skip + log<br/>in Sheets"] --> H1([End])

  subgraph AI_Validation [🧠 AI Extraction + Validation]
    B1["(5) OpenAI<br/>extract JSON summary"]
    B2{"Schema-valid<br/>JSON?"}
    B1 --> B2
    B2 -- No --> B1R["Retry<br/>stricter prompt"]
    B2 -- Yes --> C1["(6) Code step<br/>parse JSON → fields"]
  end

  A4 --> B1
  C1 --> D1["(7) Sheets<br/>append run log"]

  subgraph Routing [📊 Routing & Actions]
    D1 --> D2{"(8) Risks<br/>detected?"}
    D2 -- Yes --> E1["Slack<br/>review thread"]
    D2 -- No  --> F1["Move to<br/>/Contracts/Approved"]
    F1 --> F2["(9) Gmail<br/>confirmation email"]
  end

  E1 --> H1
  F2 --> H1

Workflow Steps

Trigger: Gmail — new attachment containing “contract”
Notify: Send immediate Slack notification (intake)
Filter: Only process contract files (skip & log others)
Store: Upload to Google Drive → /Contracts/Incoming
Extract: OpenAI step creates JSON summary (parties, dates, amounts, renewal, risks)
Parse: Code by Zapier converts JSON to typed fields
Log: Append all run details to Google Sheets
Route: If risks found → Slack human review thread; else continue
Confirm: Send email confirmations for auto-approved contracts

🎬 Try it live

Paste contract text (or use the example below) and this calls the actual deployed FastAPI backend — not a canned response.

⚠️ Not legal advice. This is a demo of an AI document-scanning tool, not a substitute for review by a licensed attorney. Do not rely on its output for real contract decisions.

📧 Support Email Agent

Type: Agentic Workflow
Stack: Zapier, OpenAI, Gmail, Zendesk, Slack, Google Sheets

This agent triages every inbound support email, drafts a reply grounded in your knowledge base, and routes anything customer-facing through a one-click Slack approval before it ever leaves the outbox.

Highlights

24/7 first-response coverage — no ticket sits untouched overnight
Human-in-the-loop approval keeps a person accountable for every reply
Confidence-based routing sends only low-risk replies (e.g., password resets) automatically
Daily Slack digest tracks response time and escalation rate

🧩 Workflow Diagram

📧 Support Email Agent — Diagram

%%{init: {'flowchart': { 'htmlLabels': true, 'wrap': true, 'nodeSpacing': 60, 'rankSpacing': 80 }}}%%
flowchart LR
  subgraph Intake [📥 Intake]
    A1["(1) Gmail/Zendesk Trigger<br/>new support email"]
    A2["(2) OpenAI<br/>classify intent + urgency"]
    A1 --> A2
  end

  A2 --> B1{"(3) Escalation<br/>keywords? refund/cancel/legal"}
  B1 -- Yes --> H1["Route to<br/>human queue"] --> H2([End])

  B1 -- No --> C1["(4) KB Lookup<br/>Sheets/Airtable match"]

  subgraph AI_Draft [🧠 Draft + Review]
    C1 --> D1["(5) OpenAI<br/>draft grounded reply"]
    D1 --> D2{"(6) Confidence<br/>+ risk check"}
    D2 -- Low-risk/high-confidence --> E1["Auto-send"]
    D2 -- Needs review --> F1["(7) Slack thread<br/>Approve / Edit / Reject"]
  end

  F1 -- Approved --> E1
  F1 -- Edited/Rejected --> D1

  E1 --> G1["(8) Send via<br/>Gmail/Zendesk API"]
  G1 --> G2["(9) Log to Sheets<br/>+ daily digest"]
  G2 --> H2

Workflow Steps

Trigger: Gmail/Zendesk — new inbound support email
Classify: OpenAI scores intent (billing, technical, general) and urgency
Escalation check: Refund/cancel/legal/angry-sentiment keywords route straight to a human, bypassing AI drafting
Retrieve: Pull the matching FAQ/knowledge-base article via lookup table
Draft: OpenAI generates a reply grounded in the matched KB content, in brand voice
Confidence check: Low-risk, high-confidence replies (e.g., password reset) go straight to send
Human-in-the-loop: Everything else posts to a Slack thread with Approve / Edit / Reject actions
Send: Approved replies go out via the Gmail/Zendesk API
Log & report: Every reply logged to Sheets; daily Slack digest shows volume, response time, escalation rate

🗒️ Meeting Intelligence & Action-Item Agent

Type: Agentic Workflow + Document Intelligence
Stack: FastAPI, OpenAI (GPT-4o-mini), Whisper, MongoDB, Slack API, Notion/Asana API

Meetings generate decisions and action items that usually die in someone’s notes app. This service listens for a finished meeting, turns the recording or transcript into structured minutes, and pushes every action item — with an owner and a due date — straight into the team’s task tool.

Highlights

Cuts meeting write-up time from ~20 minutes to under 2
Action items land in Notion/Asana automatically, no manual re-typing
Weekly rollup flags overdue items by owner, closing the accountability gap
Searchable meeting history in MongoDB — “what did we decide about X?” answered in seconds

🧩 Workflow Diagram

🗒️ Meeting Intelligence Agent — Diagram

%%{init: {'flowchart': { 'htmlLabels': true, 'wrap': true, 'nodeSpacing': 60, 'rankSpacing': 80 }}}%%
flowchart LR
  subgraph Intake [📥 Intake]
    A1["(1) Webhook<br/>Zoom/Teams recording ready"]
    A2["(2) Whisper<br/>transcribe audio"]
    A1 --> A2
  end

  A2 --> B1["(3) FastAPI<br/>send transcript to LLM"]

  subgraph AI_Extract [🧠 Extraction]
    B1 --> C1["(4) OpenAI<br/>extract summary + decisions + action items"]
    C1 --> C2{"(5) Schema-valid<br/>JSON?"}
    C2 -- No --> C1R["Retry<br/>stricter prompt"] --> C1
    C2 -- Yes --> D1["(6) Store in MongoDB<br/>meeting record"]
  end

  subgraph Action [📊 Action + Notify]
    D1 --> E1["(7) Create tasks<br/>Notion/Asana API"]
    D1 --> E2["(8) Post summary<br/>to Slack channel"]
  end

  E1 --> F1["(9) Weekly cron<br/>check overdue items"]
  F1 --> F2["Slack digest<br/>to owners + organizer"]
  E2 --> H1([End])
  F2 --> H1

Workflow Steps

Trigger: Zoom/Teams webhook fires when a recording is ready (or a transcript is uploaded manually)
Transcribe: Whisper API converts audio to text if a transcript wasn’t already provided
Send to LLM: FastAPI service forwards the transcript for structured extraction
Extract: GPT-4o-mini pulls out a summary, key decisions, and action items (task, owner, due date)
Validate: Schema check on the returned JSON; retry with a stricter prompt on malformed output
Store: Structured meeting record saved to MongoDB for searchable history
Create tasks: FastAPI calls the Notion/Asana API to create one task per action item, tagged to the right owner
Notify: Clean summary + action item list posted to the team’s Slack channel
Track & digest: Weekly cron checks the DB for items past due date and Slack-DMs each owner, with a rollup to the meeting organizer

🎬 Try it live

Paste a meeting snippet (or use the example below) and this calls the actual deployed FastAPI backend — not a canned response.

Milele Hallingquest

🧠 AI Automation & Data Engineering Projects

📊 Medicare Data Pipeline: Raw Data to Production Dashboard

🧩 Workflow Diagram

📊 Medicare Data Pipeline — Diagram

📈 Superstore Sales & Profitability Dashboard

🧩 Workflow Diagram

📈 Superstore Pipeline — Diagram

🗳️ 2026 Congressional Campaign Finance Dashboard

🧩 Workflow Diagram

🗳️ FEC Campaign Finance Pipeline — Diagram

🏀 NBA Team Performance & Shooting Trends

🧩 Workflow Diagram

🏀 NBA Team Stats Pipeline — Diagram

🎯 March Madness ML: 2025 NCAA Bracket Predictions

🧩 Workflow Diagram

🎯 March Madness ML — Diagram

⚙️ Contract Processing Bot

🧩 Workflow Diagram

⚙️ Contract Processing Bot — Diagram

📧 Support Email Agent

🧩 Workflow Diagram

📧 Support Email Agent — Diagram

🗒️ Meeting Intelligence & Action-Item Agent

🧩 Workflow Diagram

🗒️ Meeting Intelligence Agent — Diagram