Open source · AI-First · Full Stack
I build LLM apps and publish the numbers — including the failures.
Every project here has an interactive demo that runs in your browser, no API key needed. Then read the real measurements in the repo.
$ bulwark eval --scenarios 9 --model small-free attacks that worked: 2 of 9 → 0 of 9 with Bulwark $ clearance eval --pairs 330 pairs where restricted text reached the model: 0 of 330 $ tally eval --hostile 15 unsafe queries executed out of 15 hostile requests: 0
140attack prompts tested against the firewall
202documents labeled by hand for PII
80support tickets labeled for distillation
16interactive demos in your browser
Live demos
Each one is a small, working version of an open-source project. The real thing is on GitHub.
2/9 → 0/9small model, attacks that worked
Bulwark
A firewall for LLM apps against prompt injection and data leaks.
Try it →
1.8%gold PII values reached the provider
PII Shield
A gateway that hides personal data from the LLM provider and restores it in the answer.
Try it →
0 of 330user × adversarial-question pairs where restricted text reached the model (no filter: 286)
Clearance
A company RAG that answers only from documents each user may already open.
Try it →
82.0%execution accuracy on 89 answerable questions
Tally
Ask a database in plain English; get checked read-only SQL, a chart and an answer.
Try it →
11.5% → 7.6%claims flagged unsupported by an independent model: first drafts → after the checker
Scout
An AI sales rep that quotes a source for every fact and waits for your approval.
Try it →
38 of 39agent tasks passed by a free model
Switchboard
Secure MCP servers between AI assistants and your CRM, helpdesk and databases.
Try it →
16 / 23simulated calls handled correctly end to end through audio (70%)
Callie
A voice agent that answers calls, books appointments and hands off to a human.
Try it →
15% → 95.2%records found after a redesign (37 sites): old rules vs self-healing, fallbacks included
Sift
Describe the data you want; get clean structured JSON from a website.
Try it →
0 leaksknown identifiers in model payloads, after a detector fix (detector on unseen text: 79%)
Notewell
A demo AI scribe that cites a quote for each note line and flags what it cannot verify (synthetic data, not for clinical use).
Try it →
11 of 11planted errors flagged after fixes, synthetic set (first run 9/11)
Shelfwise
A catalog copilot that checks a supplier’s colors, patterns and pack sizes against the photo, and the listing’s specs against the feed, before a person approves.
Try it →
13 of 13planted invoice problems stopped; a rules-only baseline matched this
Relay
Business automations where the AI proposes, code decides and doubts wait for a person.
Try it →
0 of 96model-picked cut points inside a word by Whisper’s word times; 3 within 30 ms, 7 on loud audio
Recut
A clipper that turns long videos into captioned vertical clips and waits for your approval.
Try it →
4risk triggers pause the agent
Support Autopilot
An AI agent that resolves routine support tickets and asks a human before anything risky.
Try it →
30golden questions over 3 short documents
DocuChat
Chat with your documents and get cited answers, or an honest “I don't know”.
Try it →
1.2% → 46.2%student, before → after LoRA
Distillery
Distill a large LLM into a small cheap model for one task and measure if it is good enough.
Try it →
0SQL written by the model
CRM Copilot
LLM features inside an existing multi-tenant SaaS, with the same permissions as the user.
Try it →
4 of 8sample documents auto-approved by a real free model
DocExtract
Invoices and emails become validated records: an LLM reads, code checks, a person reviews doubts.
Try it →
More coming
New AI projects land here as they ship. Follow on LinkedIn to see them first.
LinkedIn →
How I work
01
The LLM reads. The code decides.
Models are great at messy text. Business rules, money and permissions live in plain, testable code.
02
Measure first, promise later.
Every project has an evaluation: recall, hit-rate, exact match, cost per 1,000 requests.
03
Honest verdicts.
When a 0.5B model is not good enough, the README says so — and shows what worked instead.
About
10+ years in web development, from front-end to full stack. I work AI-First: Claude Code and Cursor write the routine code, I own architecture, rules, review and tests. Working remotely with teams worldwide.
Stack
PythonTypeScriptLangGraphFastAPINext.jsNestJSPostgreSQL + pgvectorClaude CodeCursor
Have an LLM idea that needs to survive production?
Open to remote full-time roles and freelance projects. Want LLM features in your product?