▲ Kymaron Labs · Sebastian Pozzo

AI, built
to spec.

I've spent fifteen years developing machine learning and intelligent software systems, working with large language models since their earliest practical applications. At Kymaron Labs, I turn frontier AI from impressive demonstrations into precise, fast and dependable technology ready for real-world use. And I still write code myself.

DWG KL-004 · Lever movement
N = 10
error = 38.37% · 20 spinning circles
z(t) = Σ ck eikt  ·  |k| ≤ N
15yrsBuilding AI systems
780Commits a month, last 6 months
400k+Lines of code a month
4,700Commits since April
▲ Toolbox

AI extends far beyond language models and their interfaces.

LLMs are the newest tool in my kit, not the whole kit. Most of the craft is knowing when a decision table beats a prompt, and when a solver beats both. I've shipped all of these, and the best systems I build combine several of them.

Classical AI

Reasoning & search

  • Search & planning A* · MCTS
  • Constraint solving CP · SAT
  • Optimisation LP · MIP
  • Rule-based systems Rule chaining
  • Knowledge representation Knowledge graphs · ontologies
  • Probabilistic reasoning Bayesian networks
Machine learning

Learning from data

  • Tree ensembles RF · GBDT
  • Clustering & outliers DBSCAN · LOF
  • Forecasting Time series
  • Computer vision Detection · OCR
  • Reinforcement learning Policy learning
  • Recommenders Ranking
Decision & generative AI

Decisions, language & media

  • Decision models Jev
  • LLMs & agents Tool use
  • Grounding & retrieval RAG
  • Image generation Diffusion
  • Speech processing ASR · TTS
  • Generative 3D Meshes
15 years ago · classical AI & ML
Today · frontier models
▲ Recent frontier work

Where frontier models fall short, I engineer the rest.

Every frontier model is brilliant in places and unreliable in others. My work is finding those edges and building the systems that make the results dependable.

3D & simulation

Generative 3D, real physics

3D generated by AI that has to behave like the real thing.

  • Physics engines. Generated assets are handed to traditional physics engines, which keep them honest.
  • Best of both. Generative where that helps, deterministic where it has to be.
Real-time media

Live AI video & audio

Video and audio streams generated by AI in real time, as they play.

  • Generated live. Nothing is rendered in advance and played back.
  • Built for streaming. Engineered for live delivery, where every frame has a deadline.
Imaging

Precision technical schematics

Image models give you a picture of a circuit, not a circuit. I needed the real thing.

  • Spiky vision. Every frontier model writes well, but their vision is uneven. I map where it can be trusted and verify everything else another way.
  • Exact drawings. The AI drives real technical drawing tools, so diagrams come out labelled, to scale and correct.
Voice

A receptionist you'd take for a person

Speech to text and text to speech with a response lag so small the conversation feels seamless.

  • Mixed model sizes. Small, fast models keep the conversation moving while larger ones think.
  • Streaming. Techniques that overlap listening, thinking and speaking hide whatever wait is left.
Publishing

Textbooks on the fly

An exam specification goes in. A complete textbook comes out, beautifully written, illustrated and diagrammed.

  • House voice. Word-pattern and frequency analysis steers the prose away from the tells of machine writing.
  • Zero hallucinations. Every claim is grounded in the specification and its sources, and checked before it reaches the page.
Decisioning

System One models + frontier LLMs

Frontier LLMs are brilliant, but slow and expensive to call for every judgment a system makes.

  • Decide fast. System One models such as Jev answer typed questions (classify, score, yes or no) with calibrated probabilities, far faster and cheaper than an LLM.
  • Think where it counts. The frontier LLM handles only the open-ended reasoning, so results come back sooner, cost less, and every decision returns a typed, auditable signal.
▲ The lab & the pace

Hands on the code, every day.

I don't just advise from a slide deck. I'm in the codebase, and I'm always experimenting with whatever is new at the frontier. Here is every commit I've made over the last six months.

780Commits a month, averaged over 6 months
400k+Lines of code a month
1,242Commits in August, my busiest month
Commit activity · Apr – Oct 2026Busiest day · 29 Sep · 301 commits
4,700 commits · 159 active days Fewer More

Something new comes out of a frontier lab nearly every week. I test it that same week against real problems, not benchmarks, and keep what earns its place. That's how my clients end up with techniques their competitors haven't heard of yet.

▲ Who I work with

From ambitious ideas to global enterprises.

I work with anyone who wants genuinely cutting-edge solutions, from brand-new startups building their first product to large companies that need something nobody has built before.

Enterprise

Code I've written is in products adopted by FTSE companies.

Public sector

My code was the technological base of a Microsoft EU tender.

Channel

My code is the core of a product recommended by Microsoft to the UK IT channel.

Silicon

I'm the chief technology partner of a company that is an official supplier to Intel.

Ambitious idea
Global enterprise
▲ Work with me

Bring me the hard problem.

Sebastian Pozzo · Founder, Kymaron Labs

Strategic AI Advisor

Advising global technology leaders on AI strategy, technical architecture and complex engineering challenges.

Systems Architecture & Diagnostics

Diagnosing, designing and engineering complex systems, from established enterprise infrastructure to frontier AI applications.

Fractional CAIO

Serving as Fractional Chief AI Officer across multiple companies, directing AI strategy and engineering while remaining hands-on in the code.