Case Studies

LLM decision support · Enterprise

Agent Marketing Simulator

An LLM tool that evaluates emails, portal pages and designs before launch and returns a forecast, a named A/B winner, element-level fixes and risk flags. Built at Liberty Mutual; described here at the level of architecture and method only.

Problem

Pre-launch marketing reviews mixed opinion with numbers, and there was no record of whether a past prediction had been right. A model that sounds confident is not the same as one that has been checked.

What I built

Numbers and judgement are kept separate by design: a deterministic forecast layer produces every figure, and a stratified persona panel (built with Anthropic's Claude models) supplies the qualitative read. Governance rules: eight disconnected dashboard exports consolidated into one calibration baseline, a four-level sourcing hierarchy for every printed number, confidence caps, and a rule against naming a winner inside statistical noise.

Result

Validated against six live production A/B tests, with every miss logged permanently. Grading exposed two systematic model biases and one wrong segment assumption, all corrected. Post-fix results are deliberately excluded from the record because they were fit to the same tests.

• Claude
• Multi-agent personas
• Evaluation & calibration
• Data governance

Natural language to SQL · Local LLM

ESC Family History Explorer

Three legacy Microsoft Access databases turned into one SQLite database that anyone can question in plain English.

Problem

A family's expedition archive (photos, journals, people, GPS trips) lived in three legacy Access databases that were hard to search, and standard export tools could not decode some of the data.

What I built

Converted the databases to SQLite (about 144K images, 3K people, 6.9K trips) and built a FastAPI service with a local LLM that turns plain-English questions into SQL. Guardrails: read-only SQL validation, query timeouts, result caps, and schema metadata that steers the model away from known mistakes. On top of that: a photo browser, journal magazine, D3 family tree, Leaflet trip maps and a Three.js 3D terrain viewer built on USGS elevation data.

Result

Benchmarked four models; the chosen one passed the full 10-question internal test set. Recovered 2,262 historic dates that standard export tools could not decode. The system runs behind a login with identity verification, so there is no public demo.

• SQLite
• FastAPI
• Ollama
• D3
• Leaflet
• Three.js

Multi-agent automation

Bedrock: an insurer run by AI agents

A demo insurance company whose website is planned, illustrated, coded and published each day by a team of autonomous agents, working from real market data.

Problem

Explore how far a small set of specialised agents can run a content operation end to end, and where they need guardrails.

What I built

Four role-specific agents and an orchestrator, run once a day on a local Apple Silicon machine. The Director writes a brief from live market data and news; the Photo Designer renders a hero image with Flux while the Web Developer writes the page copy; the Publisher commits the change and pushes it to the running site. Every figure on the page is computed from fetched data, never written by the model, and a value that can't be fetched shows a dash. If a render fails, the previous image stays. Visitors can replay the latest meeting from the saved log.

Result

A full run takes about 90 seconds, down from several minutes that often ended in a timeout. The most useful lesson came from reading the logs: an early version silently substituted random prices when its data source blocked it, and the model then presented them as live. I removed every fallback that could invent a number and made failures visible instead.

• Multi-agent orchestration
• Local LLMs
• ComfyUI / Flux
• Data integrity
• Python

Generative media

Generative film restyling pipeline

Public-domain films repainted frame by frame in the style of a chosen artist, with a review workflow to keep characters consistent.

Frame from Beat the Devil (1953, public domain).

Problem

Off-the-shelf video models drift: faces change and details hallucinate from frame to frame. Restyling a whole film needs consistency, not just one good image.

What I built

A pipeline that replaces each frame with a repainted equivalent using Stable Diffusion 1.5 and SDXL, ControlNet lineart to hold the composition, and custom-trained LoRAs (including a Reginald Marsh SDXL LoRA trained on 74 images). A frame-review workflow catches character drift and hallucination, flagged frames are re-rendered at tuned settings, and optical-flow smoothing evens out the motion.

Result

Several public-domain films restyled end to end in different artist styles. Generative video is a separate strand of the same work: the reel on the home page was made with Gemini, Google Veo 3.1 and ElevenLabs.

• Stable Diffusion
• ControlNet
• LoRA training
• ComfyUI
• ffmpeg