I build LLM infrastructure and agentic systems in production, and I study how agents behave when they act through tools.
Sole technical lead of a 17-person engineering team at Blackwired, a threat intelligence company. Ten years across AI platforms, low-latency trading systems, and a product I founded and grew to 50,000 monthly users.
Run manifestproduction and research
$30k+/mo
cut from LLM inference costs with high-throughput serving
Blackwired
50k pages/mo
of client reports written by an autonomous intelligence agent
Blackwired
17
engineers, as the team’s sole technical lead
Blackwired
83,584
agentic rollouts, graded on what the agent did to its environment
Policy in Context
Selected work
Research · 2026
Policy in Context
Across two agentic domains and three open-weight models, one variable governed whether an agent discharged a mandated duty or obeyed an instruction its policy forbade: whether it had read that policy into context. A placebo-controlled test showed the policy’s content does the work. A same-shaped document with the duty rules removed, delivered through the same channel, did not reproduce the effect.
The most useful result came from checking the other side of the ledger. With its policy handed over and reasoning switched off, the smaller model escalated on routine incidents that needed no call at all.
159/160
reasoning off policy in context
16/160
reasoning on same policy, same model
Putting a policy in context is necessary, and on this model it is not enough. The reasoning trace supplies a second thing, judgement about when the policy applies.
Built as an agentic evaluation harness: tool-calling loops against locally served models, with grading read from the environment and no LLM judge. I directed the study and ran it with an AI coding agent as the research engineer. Five adversarial verification rounds caught errors, including errors in the corrections.
My entry to OpenAI’s challenge to train the best language model that fits in 16 MB and trains in under ten minutes on 8×H100s. An evaluation-time technique that mixes the model’s predictions with a PPM-D byte model through a high-confidence gate. It is worth −0.034 bits per byte over sliding-window evaluation alone, for a final score of 1.067219, 0.006 behind the merged record when I submitted (1.0611).
Within a day of my submitting, a reviewer found that my headline divided by the wrong byte count. I published the correction on the entry itself and kept the original logs up as the record of what ran.
“Errata … corrected headline val_bpb 1.067219 (was 0.979556).” “On the canonical denominator the corrected mean is +0.006 BPB worse than PR #1855’s merged SOTA (1.06108), so this is not a SOTA submission.”Errata on the entry, 2 May 2026
