MUHAMMAD ATASHI
Systems that make model behavior measurable, inference economically defensible, agent authority explicit, and real-world automation safe to operate.
ENGINEERING POSITION
Models are probabilistic. Production systems do not get to be vague.
I design deterministic boundaries around the parts of AI systems that carry money, authority, release risk, security impact, or external claims.
The work below is organized by the question each system owns — not by framework, model vendor, or demo format.
FLAGSHIP SYSTEMS
InferenceLedger decides which execution path is defensible. AgentGuard binds authorization to runtime execution. ProofDiff makes behavioral change reviewable before release.
Inference economics · constrained execution
Savings without acceptable quality are not savings.
An evidence-driven inference optimization and decision system that selects execution strategies under explicit cost, latency, quality, reliability, capacity, and deadline constraints.
Runtime execution integrity · agent authorization
Authorized Action == Credentialed Action == Executed Action.
A runtime execution-integrity boundary that binds deterministic authorization, least-privilege credentials, execution-time revalidation, replay protection, and terminal audit evidence.
Is this exact action safe to execute?Runtime action authorization + execution integrityOpen case study
Behavioral regression · AI change assurance
Traditional software has code diff. AI systems also need behavior diff.
A baseline-vs-candidate behavioral change-analysis system for model, prompt, retrieval, agent, tool, memory, orchestration, and decoding changes.
What changed in the behavior of my AI system, why did it change, and is it acceptable?Behavioral change assuranceOpen case study
SYSTEMS LANDSCAPE
Behavior, permission, routing, runtime authority, external claims, and business actions are deliberately separate concerns. The architecture becomes stronger when those boundaries stay explicit.
GOVERNED SYSTEMS
These systems keep high-impact decisions inspectable: release-time permissions, evidence-backed enterprise claims, and policy-gated support actions.
APPLIED / INDUSTRY SYSTEMS
Durable background work, multilingual product behavior, heterogeneous evidence, and stateful monitoring — with real interfaces instead of architecture-only diagrams.
Client / company system · end-to-end AI content operations
A production-oriented multi-stage content operations platform coordinating research, context retrieval, planning, generation, quality control, SEO, review, publishing, and operational monitoring across AI providers.
Anonymized commercial system · governed multi-channel conversational commerce
Persian-first governed conversational commerce: persona-driven consultative selling with grounded product truth, safety-before-sales enforcement, bounded commercial relationships, recent-turn context, and multi-channel delivery across web widget, Telegram, and Instagram DM.
Evidence-grounded analysis · multi-format ingestion
Domain-neutral multi-format analysis with deterministic profiling, hierarchical source-grounded synthesis, strict source/chunk validation, and adaptive Telegram reporting.
Automation · state reconciliation
Scheduled listing discovery, pagination coverage, persisted state reconciliation, detail validation, and controlled Telegram notification for a real-estate monitoring workflow.
RESEARCH ↔ ENGINEERING
CONTACT
