All projects

Every system I can show, each told as the problem, what I built and the result. Client work under NDA is marked Confidential. Source code is on GitHub.

Air-gapped exam network NO INTERNET · LOCAL SUBNET ONLY CSV question bankmessy, inconsistent Pandas ETLnormalise · validate SQLite storeclean, queryable PySide6 serverasync HTTP pollingexam distribute + collect Candidate 01 Candidate 02 Candidate n… Zero exposuremeets aerospace protocol

NASTP Examination Network

Aerospace, secure offline system

Problem
Security rules banned any internet connection, which ruled out every off-the-shelf exam platform.
What I built
A fully offline client-server system using asynchronous HTTP polling across a local subnet, with a Pandas ETL loading messy CSV question banks into SQLite behind a PySide6 desktop app.
Result
Digital exams running inside a facility where cloud software was never an option.
  • Python
  • PySide6
  • SQLite
  • Pandas
  • Sockets
Repository
Veridian pipeline Global listingsunstructured, 100s of sites Playwrightheadless harvest Pandas warehouseSQLAlchemy persist ReAct loopLlama 3.3 70BGroq · zero-framework SQL toolkit Web gateway Fin. calculator Analyst reporton demand, unattended

Veridian

PropTech, autonomous agent

Problem
Property research meant harvesting listings from dozens of incompatible sources, then rebuilding the same financial model by hand for every question.
What I built
Headless Playwright harvesting into a Pandas warehouse persisted with SQLAlchemy, plus a zero-framework ReAct loop on Groq's Llama 3.3 70B that routes between SQL tools, a live web gateway and deterministic financial calculators.
Result
Research that took an afternoon now runs unattended and answers on demand.
  • Python
  • FastAPI
  • Playwright
  • Llama 3.3
  • Groq
  • SQLAlchemy
Repository
Aisle retail agent Store dataorders · SKUs · customers WarehouseAPScheduler refresh Tool-calling loopGroq · Llama 312 registered tools Forecast Pricing RFM segments Next.js dashboardtoken-by-token streamanswers in seconds

Aisle

E-commerce, multi-tool agent

Problem
Pricing and demand calls were made from stale spreadsheets, because a real answer meant booking an analyst's afternoon.
What I built
An AI analyst that consolidates store data into one warehouse, with a tool-calling agent on Groq and Llama 3 using twelve tools to forecast demand, recommend prices and segment customers, streaming into a Next.js dashboard.
Result
Decisions backed by live data, answered in seconds.
  • Python
  • LangChain
  • Groq
  • Llama 3
  • Next.js
Repository
SSRF-hardened extraction "Get me X from Y"plain English LangGraph plannerruntime schema gen SSRF guardblocks loopback, private,cloud-metadata ranges Two-stage fetchrenders only when needed Pydanticschema validate Structured JSONany site, no bespoke code

AI Web Scraping Agent

Security, autonomous extraction

Problem
Every new data source meant another hand-written scraper to maintain, and naive fetchers are an open door to SSRF.
What I built
An agent that turns plain-English requests into schema-validated JSON. A two-stage fetch keeps token costs down, schemas are generated at runtime, and every resolved IP is checked against private and cloud-metadata ranges.
Result
One agent replaces a growing library of brittle site-specific scrapers.
  • Python
  • LangGraph
  • Pydantic
  • SSRF hardening
Repository
Continuous price tracking Competitor sitesscheduled sweep Requestsextract Pandas typing gatemalformed rows never stored SQLite · SQLAlchemyrelational history FastAPIsecured endpoints Chart.js dashboardlive comparisonno more manual checks

Competitor Price Tracker

Data engineering, automation

Problem
Competitor prices were checked by hand, so they were checked rarely, and the data was stale by the time anyone acted.
What I built
A scheduled pipeline that extracts prices, enforces strict typing in Pandas so bad rows never reach storage, loads SQLite through SQLAlchemy and serves a Chart.js dashboard over FastAPI.
Result
A recurring manual chore became a continuous, always-current feed.
  • Python
  • FastAPI
  • Pandas
  • SQLAlchemy
  • Chart.js
Repository
Three sources, one warehouse REST API BeautifulSoup CSV exports no shared format Pandas reconcileclean · dedupe · unify PostgreSQLsingle warehouse FastAPIserve Tailwind dashboardone place to lookno reconciliation by hand

MarketSync ETL

Data engineering, warehousing

Problem
Job market data lived in three systems with no common format, so every report started with manual reconciling.
What I built
Extraction from REST APIs, BeautifulSoup scrapers and CSV exports, reconciled in Pandas into one PostgreSQL warehouse and served through FastAPI to a Tailwind dashboard.
Result
One dashboard replaced three incompatible sources and the manual work between them.
  • Python
  • BeautifulSoup
  • Pandas
  • PostgreSQL
  • FastAPI
Repository
Non-blocking ingestion Client uploadlarge CSV OAuth2 / JWTauth gate FastAPIdomain-driven layout 202 Accepteduser never waits BackgroundTasksPandas · off the request thread DatabaseAlembic migrations high concurrency · heavy work never blocks the product

FastAPI Data Ingestion Engine

Backend, data as a service

Problem
Large uploads tied up request threads, so the product stalled for everyone whenever one customer sent a big file.
What I built
A FastAPI backend with OAuth2 and JWT, Alembic migrations and a domain-driven structure. BackgroundTasks with Pandas absorb large CSV jobs without blocking requests.
Result
Instant responses under load, with heavy processing finishing in the background.
  • Python
  • FastAPI
  • OAuth2 / JWT
  • Alembic
  • Pandas
Repository
Kaleo agent Website chat WhatsApp Email & documents Kaleo agentreads, decides, actsday and night Calendar Price list Records Job bookedquoted, confirmed, logged

Kaleo

My company

Problem
Small businesses lose work when messages go unanswered and quotes, bookings and paperwork pile up.
What I built
AI agents that reply on the website, WhatsApp and email, quote from the real price list, book jobs into the calendar and keep records and documents in order.
Result
Customers get an answer in seconds, at any hour. The owner only handles what needs a person.
  • Python
  • LLM agents
  • Tool calling
  • Calendar APIs
Visit kaleo.cc
Confidential

Agent orchestration engine

Confidential client, data operations

Problem
The client's data processes needed an agent that could reason through several steps, call outside APIs and run backend jobs without a person in the loop. A chatbot wasn't enough.
What I built
A custom agent in Python with its own tools for pulling data, calling web endpoints and writing to the database. It runs behind a FastAPI backend, and every automated action is logged.
Result
Delivered and paid on time, and the client hired me again for the next build.
  • Python
  • LLM APIs
  • Function calling
  • FastAPI
  • PostgreSQL
Under NDA. Details on a call.
Confidential

Production RAG agent

Confidential client, returning

Problem
They needed a production AI agent that answers from their own documents and returns structured, checked output.
What I built
A LangChain agent with document search (RAG), tool calling, a FastAPI backend, output checks, and full setup and handover docs.
Result
Taken from requirements to tested delivery in two days.
  • Python
  • LangChain
  • RAG
  • FastAPI
  • Pydantic
Under NDA. Details on a call.
Confidential

Data models and ETL

Confidential company, two-month engagement

Problem
The team needed reliable data models and pipelines for their projects.
What I built
Worked with their team for two months at up to 20 hours a week. I analysed their data needs and built the data models and ETL processes.
Result
A two-month ongoing engagement, not a one-off task.
  • Python
  • ETL
  • SQL
  • Data modelling
Under NDA. Details on a call.
Confidential

One-click PDF data extraction

Confidential client, document automation

Problem
Nine fields had to be copied out of every PDF by hand.
What I built
A button in Excel runs a Python script that reads every PDF in a folder and pulls out the nine fields. It writes them into the same sheet, then renames and moves each file by the company name it found.
Result
Copying data out by hand replaced with one click.
  • Python
  • PDF parsing
  • Excel VBA
Under NDA. Details on a call.

Have something like this to build?

Book a call