GigiBeridze
One developer.A team of agents.
I'm Gigi, a developer from Georgia. I build AI agents reliable enough to trust with real work.
Most of my time goes into agentic systems: giving models good tools and splitting work across agents. I build for clients on Freelancer.com, Upwork and Fiverr, and since June 2025 I've been working on Gigr's Android app and its AWS backend.
Alongside that, I'm learning machine learning from the maths up.
Open for new projects
- Rating
- 5.0 on Freelancer.com, from 5 reviews
- Experience
- 7+ years writing code
- Languages
- English and Georgian (native)
Four roles.One system.
An agent is only as good as the app around it and the infrastructure under it. I work on every layer, so they're designed to fit together.
Agentic systems: orchestrators that plan, sub-agents with real tools, and evals that prove it works before it reaches users.
Claude APIOpenAI APITool useMCPMulti-agentRAGEvalsPython
Things I've built.
From agent systems and ML experiments to production apps and cloud backends. Open any of them for the details.
A custom harness where an orchestrator plans the job, hands tasks to specialist sub-agents with their own tools, and reviews the results before anything ships.
Claude APITool useMCPSub-agents
Reliable beats impressive.
A demo that works once is easy. This countdown is how an agent gets good enough to run without someone watching it.
T−3
Scope the job
Pin down exactly what the agent does, and what “done” looks like, before writing a prompt.
T−2
Give it tools
Small, well-described tools and focused sub-agents instead of one giant prompt that does everything.
T−1
Measure it
Evals on real cases, run on every change, so it’s tested before your users are the test.
Liftoff
Ship and watch
Deployed on infrastructure you own, with logs and traces from day one so nothing fails quietly.
Same model.Better results.
Strictly for fun: Claude Opus 5.5 on its own, then the same model with me running the job. The code is a tie. Everything around it isn't.
Claude Opus 5.5
27
vs
Gigi + Claude Opus 5.5
98
Real-Work Index, the average of 7 evaluations
Code-bench VerifiedDraw
Writes the code
MakeItPop
Turns “make it pop” into a spec
ScopeCreep 2.0
Survives “just one small change”
MondayCall
Joins your Monday call
ShipIt Diamond
Gets it live on AWS, not just written
3AM-bench Pro
Picks up when your site breaks at 3 a.m.
NoBench
Says no to a bad idea, politely
Output Speed
Lines of code per second. I read mine before they ship.
SleepBench
Hours per night. Lower wins, so Opus takes it.
Strictly for fun: every evaluation was designed, run and graded by the winner. Not affiliated with Artificial Analysis or Anthropic.
Got a job for an agent?
Tell me what you want to build or automate. I'll come back with how I'd approach it, and what it would take.