01
Trigger
A row, a message, a request
AI Reliability & LLM Evaluation Engineer · Full-Stack
I build the automation, then I build the check that keeps it honest.
I ship RAG and multi-model AI products — and the evaluation harnesses that prove they actually work. LLM-as-judge, DeepEval, stronger-model vetting.
Next.js · TypeScript · Python · RAG · LLM-as-judge · DeepEval · GSAP
Positioning
Most people stop when the AI works. I treat that as halfway.
Anyone can wire a model into a product and demo something that looks right. The hard part is the part nobody sees: knowing it's still right tomorrow, on inputs you didn't anticipate, when the model drifts.
So I pair every automation with an evaluation layer — LLM-as-judge, DeepEval, stronger-model vetting — that scores the output before a user ever sees it. The automation is the easy half. The check is what makes it honest.
Featured case study · Reliability-forward
CreatorChat
An embeddable AI chat & search widget that answers from a creator's own content — with an evaluation layer that catches hallucination before users do.
The problem
A chat widget that confidently makes things up is worse than no widget. The risk wasn't building the bot — it was hallucination and drift: answers that sound right but aren't, degrading silently as content and models change.
The build
RAG over the creator's corpus, routed across multiple providers (Claude, GPT, Gemini) so the system isn't hostage to one model's failure modes. Shipped as an embeddable widget plus browser extensions, with some infrastructure on Cloudflare Workers for low-latency edge delivery.
The check
The lead signal: an evaluation harness using LLM-as-judge plus DeepEval that scores retrieved-grounding and answer faithfulness on every change. A stronger model vets responses against the source content, so a regression is caught in evaluation — not in production by a user.
Decisions I owned
- Multi-provider by design — no single model is a single point of failure.
- Evaluation is a build gate, not an afterthought: faithfulness is measured, not assumed.
- Edge infra (Cloudflare Workers) for latency where it's on the user's critical path.
- Widget + extension delivery so it embeds into existing creator surfaces with minimal friction.
Featured case study · The thesis in miniature
Content-automation pipeline
A hands-off content pipeline that drafts, vets, and publishes — with a stronger-model gate standing between “generated” and “live.”
The problem
Auto-publishing AI drafts straight to a live site is how you wake up to embarrassing or wrong posts. Volume is easy; trust is the bottleneck.
The build
Make (Integromat) orchestrates the flow off a Google Sheet: ChatGPT produces the draft, then a stronger model vets it against the brief before anything ships. Approved drafts auto-publish to WordPress.
The check
The vetting gate is the whole point — a stronger model reviews each draft and only passing content is published. Same spine as CreatorChat: the generation is the easy half; the check is what makes it safe to automate.
The pipeline · draft → vetting gate → publish
- Stage 1 / 4
Sheet trigger
- Stage 2 / 4
ChatGPT draft
- Stage 3 / 4
The check
Stronger-model vetting gate
- Stage 4 / 4
Auto-publish to WordPress
Selected work
Things I've built.
Multi-model face recognition
A tool that runs multiple face-recognition models against the same input and compares their verdicts — DeepFace / InsightFace on Colab.
Company project - private code; walkthrough available.The eval instinct applied to vision: don't trust one model's match, cross-check several.
- Python
- DeepFace
- InsightFace
- Google Colab
audio-lab (opens in a new tab)
A browser DAW backed by Python DSP with AI-assisted control — make and shape sound in the browser.
- TypeScript
- Web Audio
- Python
- DSP
- AI
just-us (opens in a new tab)
A private space for two — WebRTC video plus shared games.
- WebRTC
- Next.js
- TypeScript
family-tree (opens in a new tab)
An interactive family tree with AI image generation — built on Next.js + Supabase.
- Next.js
- Supabase
- AI image gen
Stack
What I work with.
AI Reliability & Evaluation
- LLM-as-judge
- DeepEval
- Multi-model vetting
- RAG QA
- Faithfulness & drift checks
AI Workflow & Automation
- Agentic / RAG pipelines
- Multi-provider routing
- Human-in-the-loop gates
- Eval-gated automation
- Event-driven automation
- Webhooks & API integration
- Make (no-code orchestration)
- Auto-publishing pipelines
- Cloudflare Workers
Languages
- TypeScript
- Python
Frameworks & Runtime
- React
- Next.js
- Node
- FastAPI
Data & Storage
- Supabase
- Postgres
- MongoDB
AI / ML
- Claude
- GPT
- Gemini
- RAG
- DeepFace
- InsightFace
Platforms & CMS
- Shopify
- WordPress
How I build
Automation — with the check wired in.
Every pipeline I ship runs the same shape: a trigger, an AI step that does the work, a stronger-model check that grades it — and only verified output ships. The check is the part most people skip.
- 02
AI drafts
RAG retrieve + LLM generate
- 03
The check
Stronger model · LLM-as-judge · DeepEval
- 04
Ship
Only verified output reaches users
Output that fails the check routes back — never to users.
Contact
Hiring for reliability, or need someone to build it? Either way — let's talk.
Whether you're a team that wants AI you can trust in production, or you found me cold and have a problem worth solving — the fastest path is email.