ls -lht --time-style=long-iso ./posts/
total 58K
- -rw-r--r-- 1 kerim staff 30K context-engineering-is-half-the-harness.md
- -rw-r--r-- 1 kerim staff 28K glasswing.md
./dive --reef
Hey, I'm Kerim.
I am troubled by people prematurely accepting a smaller set of possibilities for their lives.
I built Dria, an applied AI lab where we crafted Kai, the first proactive AI software engineer. Kai found critical vulnerabilities in Apple, Coinbase, and PostHog, improved kernel performance on Nvidia, and ranked state of the art above Claude and GPT on OpenAI's security benchmark and Berkeley's GSO benchmark.
Alongside wonderful teammates, we built state-of-the-art agents and frameworks across memory, tool use, and evolutionary coding, and operated one of the world's largest crowdsourced P2P inference platforms, serving over 100,000 daily active users.
Before Dria, I co-founded Clemta, at the intersection of fintech and legaltech, scaling it to 200 countries with 15,000 businesses running on the platform. And back in 2019, an AI agency where we built ML products on GPT-2 embeddings. We had a lot of fun.
These days, I’m drawn to long-horizon agents, solving the most critical surface areas of the web before ASI falls into the hands of adversaries, and building real-life experiences that help humanity overcome its struggles, feel lighter, and appreciate the beauty of life.
cat ./shipped/README.md
Selected work
-
Mem Agent A small model that gives LLMs persistent, human-readable memory through Obsidian-style markdown.
-
Kai Security A recursive language model swarm where LLMs write code that launches other LLMs to analyze, exploit, critique, and patch a codebase, and a bug only counts once a working proof-of-concept fires.
-
KaiEvolve An AlphaEvolve-style lab where MAP-Elites islands mutate whole programs, a Thompson-sampled model ensemble learns which LLM earns its tokens, and a compounding lab notebook lets the search remember what it discovered.
-
Dria Agent Alpha A compact tool-use agent that reasons and calls tools by writing Python, not JSON.
-
Dnet, Distributed LLM Inference for Apple Silicon Clusters Running large models across networked Macs by sharding them over Apple Silicon.
-
P2P Inference Network
The beauty of being human
I'm endlessly grateful to John Coltrane, Ingmar Bergman, Esbjörn Svensson, Wong Kar-wai, Jaco Pastorius, Andrei Tarkovsky, Roy Hargrove, Pedro Almodóvar, Stevie Wonder, Paul Thomas Anderson, Brad Mehldau, and Antônio Carlos Jobim.
The beauty of feeling alive
Surfing, diving, and breathwork.
The beauty of sharing
Family, childhood friendships, and cinematography. Different lives, different struggles, all finding common ground where life feels nourished.