dap — drop your card before you arrive

@serhan

Serhan Yilmaz

yılmaz means perseverant in turkish. i tend to live up to it.

building: magnos — the ai operating system for financial intelligence. agentic palantir × mckinsey for finance.
ask: people replacing entrenched saas with ai-native systems. anyone thinking hard about agent loops, compound ai, and evals at scale. ICLR/NeurIPS folks coming through SF.
offer: if you're shipping multi-agent systems in production — i've shipped four, broken most. happy to compare notes on what worked and what was a trap. also: anything turn-taking / voice-ai / streaming inference (built streamingVAP on Furhat humanoid robots at KTH).

SS26 SFflying in from stockholm

·co-authored INCLUDE — ICLR 2025 Spotlight (multilingual eval, 44 langs, 197K QA)
·founded kAi Sabancı (600+ members, NVIDIA Student Network founding cohort, first outside the US)
·EPFL Project Charisius — 2200% perf gain on gradient aggregators, 96% runtime reduction
·KTH StreamingVAP — 93% turn-taking accuracy on Furhat humanoid robots
·CMU CASCADES — >60% over-baseline question generation

cold-intro generator —

your namewhat you're building

› how this works

four-stage compound system. one model per role. revise loop on judge failure.

A.extract · haiku 4.5 reads your input and emits structured interests + level (zod-validated)
B.retrieve · BM25 over serhan's verified bio corpus → top-k chunks. (embeddings overkill at 19 chunks; rerank is a v2 if quality stalls.)
C.draft · sonnet 4.6, prompt-cached voice samples, grounded only in retrieved facts. token-streamed.
D.judge · sonnet 4.6 with an explicit rubric — names a SPECIFIC shared artifact? avoids platitudes? no hallucinations? one revise pass on fail.

join —