AVAILABLE FOR NEW PROJECTS
AI that survives production

Your AI worked in the demo. Now it's live and breaking.

I find what's wrong, fix it, and keep it running.

Get a reliability audit See the work
Live in production across voice, messaging, and on-chain
support-agent · live
4Production systems live
0On-chain incidents
68/68Contract tests passing
<1 dayTypical reply time
The problem

AI in production breaks the same few ways.

The model is rarely the issue — the seams between systems are. This is what I get called in to fix.

It hallucinates

Invents an answer the moment it hits something it doesn't know.

It forgets

Ask "is it delayed?" and it no longer knows what you meant two messages ago.

Tools return junk

An API fails and the agent improvises around it instead of stopping.

No human handoff

Out-of-scope cases get a confident wrong answer instead of a person.

Cost blows up

One retry loop or one abuser burns through the token budget.

It breaks on restart

State doesn't survive a redeploy, so every update makes it a little worse.

What I do

Fixed-scope builds with public prices, same as my Upwork catalog.

01 · Support

AI customer support agent

from $49
  • Answers customer messages in your brand voice, around the clock
  • Hands off to a human when it should, with a reason
  • Guardrails so it stops inventing answers
02 · Voice

Custom voice AI agent

from $49
  • Real speech in and out, not a phone tree
  • Tool calling so it can get things done
  • Won't invent an answer it doesn't have
03 · Integrate

Custom MCP server

from $120
  • Connects Claude to your tools and data
  • Auth and multi-tool setups when you need them
  • Production deployment, documented
04 · Optimize

LLM cost-cut migration

from $79
  • A cost audit of where your OpenAI bill actually goes
  • Only calls that pass a quality eval move to Qwen or DeepSeek
  • Before-and-after report in real dollars
Selected work

Systems where a wrong answer costs money.

Voice AI · live

Voice assistant for air travelers

Travelers talk to it to check a flight, find a gate, or sort a missed connection — in English and Turkish, on live flight data. Guardrails keep it from inventing a flight number it doesn't have.

Delivered and live. Adversarially tested across topic switches and emergencies.
Conversational AI · production

High-volume customer messaging

Built for a client handling hundreds of inbound messages a day. Intent classification splits real questions from noise, long-term memory remembers every customer, and low-confidence cases escalate to a human.

One operator now handles what would otherwise need a full support team.
Customer service · delivered

Telegram to helpdesk pipeline

A Spanish-language support bot that routes customer DMs into one Crisp inbox, with two-way sync so replies post back to Telegram.

Delivered, paid, closed. Questions stopped falling through the cracks.
On-chain · live, audited

P2P escrow smart contract

A peer-to-peer escrow protocol on BSC Mainnet: ~2,000 lines of Solidity, 68/68 tests passing, and an independent audit with all 8 findings resolved before launch.

Live, securing real value, zero incidents.
How I work

Same four steps, every project.

Diagnose first

I find the real failure modes before quoting a fix.

Scope honestly

Clear price, clear timeline. If I haven't shipped it before, I say so.

Build with guardrails

Designed around the failure modes from day one.

Hand off clean

Documented code and a runbook your team can run without me.

Questions

The things clients usually ask.

What if my AI can't actually be fixed?

Then the audit says so, and you keep the findings. Sometimes the right answer is a different architecture.

How fast can you start?

An audit usually turns around in days. Fixes are scoped from there.

Fixed price or hourly?

Either. Small fixes work best fixed-price; larger work runs hourly or in milestones.

What's your stack?

Claude and GPT APIs, RAG, MCP, multi-agent orchestration, voice agents, Python and FastAPI, Next.js, vector databases, and the webhook and integration plumbing that ties it together.

Do you build from scratch too?

Yes. The reliability work is the same either way.

How I think about it

I build for the day after launch.

I ship AI that runs in production, handling real customers and real transactions. You get documented code, a runbook, and honest scoping.

Claude / GPT APIRAGMCPMulti-agentVoice agentsPython · FastAPINext.jsVector DBWebhooks & integrationsSolidity
Get in touch

Got an AI that's breaking in production?

Tell me what's going wrong. I'll tell you what it'll take to fix, and whether it's worth it.

Email me →Résumé

Usually replies within a day · Taipei (GMT+8)