Mikita DaroshkinMikita Daroshkin, senior AI/ML engineer, forward-deployed. Agentic AI that Fortune 500 companies run in production under audit: pharma, healthcare, lending.https://mikitadaroshkin.com/A confidence score is not a reasonhttps://mikitadaroshkin.com/posts/a-confidence-score-is-not-a-reason/https://mikitadaroshkin.com/posts/a-confidence-score-is-not-a-reason/An extraction confidence is not an adverse-action reason. Document Intelligence returns two scores, and the gate reads the wrong one.Tue, 21 Jul 2026 16:00:00 GMTThe permission model was built for people holding jobshttps://mikitadaroshkin.com/posts/agents-inside-a-system-of-record/https://mikitadaroshkin.com/posts/agents-inside-a-system-of-record/An HCM grants permissions to people holding jobs, and an agent is neither. Least privilege with no native expression, and the primitive that just shipped.Tue, 07 Jul 2026 16:00:00 GMTNobody owns the join between two agent platformshttps://mikitadaroshkin.com/posts/federating-agents-across-vendor-platforms/https://mikitadaroshkin.com/posts/federating-agents-across-vendor-platforms/Four vendor agent platforms in one mesh, each sound alone. The seven joins that failed, starting with three catalogs where the loosest one decides.Tue, 23 Jun 2026 16:00:00 GMTThe LLM judge is a model toohttps://mikitadaroshkin.com/posts/llm-judge-is-a-model-too/https://mikitadaroshkin.com/posts/llm-judge-is-a-model-too/The judge is a model too, with the same length, position and self-preference biases. Measuring it, versioning it, and knowing when not to use one.Tue, 16 Jun 2026 16:00:00 GMTContainment is measured on the calls that never needed youhttps://mikitadaroshkin.com/posts/when-the-agent-spends-the-customers-money/https://mikitadaroshkin.com/posts/when-the-agent-spends-the-customers-money/Containment counts the sessions that never needed you. A booking agent's rollback is a refund, and a surge multiple is a scalar over a rotating mixture.Tue, 09 Jun 2026 16:00:00 GMTThroughput per dollar, once the traffic is burstyhttps://mikitadaroshkin.com/posts/throughput-per-dollar-serving-under-bursty-traffic/https://mikitadaroshkin.com/posts/throughput-per-dollar-serving-under-bursty-traffic/One roofline number sets the speculative-decoding gate, the batch-size knee and the chunked-prefill floor. Why max-num-seqs is a latency control.Tue, 26 May 2026 16:00:00 GMTValidating a system that never returns the same answer twicehttps://mikitadaroshkin.com/posts/validation-as-code-under-gxp/https://mikitadaroshkin.com/posts/validation-as-code-under-gxp/Validating a system that never returns the same answer twice: what qualification becomes when the output is distributional, and where sampling breaks.Tue, 05 May 2026 16:00:00 GMTThe guarantee is marginal and the audit is nothttps://mikitadaroshkin.com/posts/calibrated-risk-when-a-person-is-the-outcome/https://mikitadaroshkin.com/posts/calibrated-risk-when-a-person-is-the-outcome/A conformal coverage guarantee is marginal, not conditional, so it can hold overall and fail on exactly the subgroup your fairness audit is about.Tue, 14 Apr 2026 16:00:00 GMTThe wall an 8B model hits on a mid-tier Android phonehttps://mikitadaroshkin.com/posts/shipping-an-8b-model-offline-on-mid-tier-android/https://mikitadaroshkin.com/posts/shipping-an-8b-model-offline-on-mid-tier-android/An 8B model on a $250 Android phone hits a bandwidth ceiling, a residency ceiling and a thermal ceiling, in that order, and dies mid-sentence.Tue, 24 Mar 2026 16:00:00 GMTTail sampling has three ways to lose the trace you wantedhttps://mikitadaroshkin.com/posts/opentelemetry-span-design-for-agents/https://mikitadaroshkin.com/posts/opentelemetry-span-design-for-agents/Tail sampling fixes head sampling and adds three quieter ways to lose the same traces, none visible in the config and none logged as an error.Tue, 17 Feb 2026 16:00:00 GMTModel Armor and DLP have to get through the perimeter toohttps://mikitadaroshkin.com/posts/agent-tools-inside-an-egress-locked-perimeter/https://mikitadaroshkin.com/posts/agent-tools-inside-an-egress-locked-perimeter/The content screens you put inline are themselves services behind the same wall. The endpoint Model Armor needs, and the field that says a screen ran.Tue, 13 Jan 2026 16:00:00 GMTWho runs the OAuth flow decides your network topologyhttps://mikitadaroshkin.com/posts/identity-delegation-across-agent-hops/https://mikitadaroshkin.com/posts/identity-delegation-across-agent-hops/Who runs the OAuth flow decides your network topology. Token exchange across agent hops, and why an act claim records who acted but never when.Tue, 11 Nov 2025 16:00:00 GMTThe handoff is the part nobody writes downhttps://mikitadaroshkin.com/posts/the-discipline-of-leaving/https://mikitadaroshkin.com/posts/the-discipline-of-leaving/Leaving a fleet of LLM agents a client's own team can run: discovery by directory, scorecards a stranger can read, runbooks drilled against real failures.Tue, 09 Sep 2025 16:00:00 GMTFaithful to the wrong versionhttps://mikitadaroshkin.com/posts/rag-on-a-regulated-corpus/https://mikitadaroshkin.com/posts/rag-on-a-regulated-corpus/On a versioned corpus, faithfulness metrics pass an answer grounded in a revision retired months ago. Treating the corpus as bitemporal is what fixes it.Tue, 17 Jun 2025 16:00:00 GMTText-to-SQL that can't leak data it shouldn'thttps://mikitadaroshkin.com/posts/text-to-sql-that-cant-leak/https://mikitadaroshkin.com/posts/text-to-sql-that-cant-leak/The seam between what a model may see and what it may run, enforced by different systems owned by different teams, and the error message as a leak channel.Tue, 18 Mar 2025 16:00:00 GMT