AI Infrastructure · Systems Engineering

I build the infrastructure around AI that has to hold when reality gets messy.

LLM systems, agent security, behavioral release assurance, inference-change decisions, and governed automation — designed around failure states, not just successful demos.

Selected work Systems atlas GitHub

In production at Setareh Hamoon, I cut inference p95 latency from 850 ms to under 200 ms — vLLM batching, Redis semantic caching, and Celery/Kafka pipelines, measured on the serving stack, not a demo.

Ten systems, five control boundaries.

Every system in the record, with the boundary it owns and the source status it actually has. The ordering below is the register, not a ranking.

NºSystemOwnsTypeSource

01AgentGuardAuthorizes tool execution and enforces runtime boundaries fail-closedFlagshipPrivate source 02InferenceLedgerAudits model and provider migrations against incomplete evidenceFlagshipPrivate source 03ProofDiffVerifies behavioral regressions against protected contractsFlagshipPrivate source 04PermitDiffV0.1.0RC2Detects authority expansion in agent permission changesLabPrivate source 05TrustFlowV0.1.0RC2Binds every external claim to approved, current evidenceLabPublic source 06ResolveOpsV0.1.0RC2Carries ambiguous external effects durably until provider reconciliationLabPublic source 07Content Automation PipelineFences publishing to the exact approved revisionProductionClient-sanitized 08Persian Conversational CommerceGrounds conversational selling in approved catalog facts and policyProductionClient-sanitized 09Universal AI AnalyzerV2.1.2Grounds multi-format extraction and hierarchical synthesisProductionPrivate source 10Centris Property MonitoringV8.1Reconciles full listing sets before any notification shipsProductionPrivate source

Runtime security

authorization · execution integrity

Release assurance

behavioral change · regression gates

Inference systems

quality · reliability · latency · cost

Applied AI ops

durable workflows · controlled effects

01—03

Three distinct failure boundaries

Three places AI can lose control.

What was allowed to execute. What changed before release. What evidence is sufficient to accept an inference change.

Execution integrity

AgentGuard

The action that executes should still be the action that was authorized.

Deterministic runtime authorization for actions routed through AgentGuard, binding authorized intent to bounded credentials, guarded dispatch, and reconciliation of uncertain effects. The linked reference, package 0.1.0, is a separate minimal example and not a newer Core release.

View case study Public Reference

AUTHORIZED INTENT / LIVE ACTION

DRIFT → REFUSE

01intentrefund.create · ≤ $100BOUND

02credentialintent fingerprintBOUND

03live callrefund.create · $150CHANGED

04dispatchside effectREFUSED

02Release assurance

ProofDiff

A change should trigger the evidence its blast radius actually requires.

Evidence-first release assurance that identifies what changed, selects the protected behaviors implicated by that change, replays the required evidence, and returns a scoped PASS / REVIEW / BLOCK decision.

Public proof point

tool:refund.create → 4 / 9 → fixed → PASS

View case study Public Reference

CHANGE

tool:refund.create

SELECT

4 / 9 contracts

REPLAY

fixed

DECIDE

PASS

Unknown impact widens the evaluation surface. It never earns a cheaper test plan.

fail closed

03Inference decision

InferenceLedger

A cheaper model is not a better migration if a required floor breaks.

Evidence-driven inference-change assurance for one model, provider, or execution-policy change, judged against requirements declared before the comparison.

View case study Public Reference

InferenceLedger retained evidence board. Two outcomes are INCONCLUSIVE, one is VALIDATED_MISMATCH, the external pilot is not started, and no retained snapshot outcome reached APPROVE. — Retained evidence boardNo snapshot reached APPROVE

04 · Applied systemoperating workflow

Content Automation Pipeline

A full-stack AI content operations platform spanning research, retrieval, generation, editorial review, revision-bound publishing, Search Console feedback, and controlled release operations.

Research, generation, review, exact-revision approval, publishing, feedback, and release operations share one explicit state model instead of leaking authority between steps.

View case study Public Reference

01RESEARCHsource-bound

02REVISIONimmutable

03APPROVALexact-revision

04PUBLISHcontrolled

05—09Second shelf

Different domains. The same insistence on explicit boundaries.

These systems extend the thesis without competing for flagship hierarchy.

Open the system atlas

01ResolveOpsCarries ambiguous external effects durably until provider reconciliationOPEN ↗02PermitDiffDetects authority expansion in agent permission changesOPEN ↗03TrustFlowBinds every external claim to approved, current evidenceOPEN ↗04Persian Conversational CommerceGrounds conversational selling in approved catalog facts and policyOPEN ↗05Universal AI AnalyzerGrounds multi-format extraction and hierarchical synthesisOPEN ↗

Research state

UNRESOLVED

is a valid state

Claims do not become stronger because the portfolio wants a cleaner ending.

External validation, inconclusive outcomes, and missing evidence stay visible instead of being converted into marketing certainty.

Validation register

Work / research / collaborationcontact

If the system needs to stay explainable when evidence is incomplete, authority changes, or reality becomes ambiguous, that is the kind of engineering problem I want to work on.

mohammadaliatashi@icloud.com