I build production-oriented AI systems with an emphasis on verification, deterministic controls, auditability, cost discipline, and explicit safety boundaries.
My work sits at the intersection of AI systems engineering, product architecture, and applied AI.
I am interested in a specific class of engineering problem:
How do we make AI useful in production without giving it authority it should not have?
That question shapes the systems I build: LLMs may reason, generate, classify, or assist, while deterministic logic, verified data, policy boundaries, observability, and human escalation define where AI is allowed to act.
| Project | Engineering Problem | Core Ideas |
|---|---|---|
| SaaSForge | Turn product ideas into feasibility-checked software scaffolds | evidence-driven evaluation · GO / NO_GO / RISKY gating · async generation · partial-failure isolation |
| AgentForge | Turn successful agent executions into reusable workflows | ReAct execution · trace capture · deterministic replay · guided replay · MCP · auditability |
| AgentCraft Marketplace | Package software capabilities as atomic AI services | service extraction · semantic combination · reusable bundles · SLA-aware delivery · credit pricing |
Discover → Evaluate → Recipe → Feasibility → Scaffold
A five-stage product-generation pipeline that researches open-source building blocks, evaluates SaaS potential, combines compatible tools, applies an explicit feasibility gate, and generates structured project scaffolds.
A controlled run produced a 19-file, ~6,400-line scaffold with developer handoff material. I classify this as VERIFIED IN A CONTROLLED RUN, not as a generalized benchmark.
Agent Run → Execution Trace → Parameterized Skill → Replay
An agent execution engine built around a ReAct tool-calling loop. Successful runs can be captured as reusable Skills. Deterministic Replay removes repeated LLM reasoning where a recorded workflow is sufficient; Guided Replay resolves runtime variables cheapest-first before escalating when necessary.
A service-oriented layer for turning software capabilities into reusable commercial AI units. Engineering evidence explicitly separates implemented, scaffolded, and planned capabilities.
| Project | Engineering Problem | Core Ideas |
|---|---|---|
| Hossaty (Tutor) | Build an AI-native operating system for private tutors | RAG · pgvector · vision grading · worker pipelines · closed-loop weakness detection |
| AgentCraft Scout | Move from company discovery to actionable B2B intelligence | enrichment · deterministic scoring · opportunity detection · multilingual outreach · CRM workflows |
| Migration Agent | Match people to immigration opportunities without allowing the model to invent programs | verify-first architecture · official-source validation · semantic matching · deterministic eligibility |
-AI%20Education%20Platform-5F7A72?style=flat-square){kind=icon}
Curriculum + Retrieved Context + Student Level + Known Weaknesses + Conversation Memory
The tutor is deliberately context-constrained rather than an unrestricted chatbot. Scheduling, payments, homework, performance tracking, and AI tutoring are treated as parts of one operational loop.
Discovery → Enrichment → Scoring → Opportunity Detection → Outreach → Deals
A B2B intelligence workflow connecting discovery, enrichment, deterministic commercial scoring, opportunity detection, AI-assisted outreach, scheduling, and CRM operations.
A verify-first architecture where AI understanding comes after source verification and deterministic eligibility logic. Programs are tied to official sources; automated submission to government portals is intentionally outside the system boundary.
| Project | Engineering Problem | Core Ideas |
|---|---|---|
| Senior Care Agent | Support elderly care without replacing family or clinical judgment | companion AI · health workflows · adherence · escalation · privacy |
| Blind Care Agent | Build useful voice-first assistance without overwhelming the user | accessibility · safety prioritization · Smart Silence · emergency support |
| Rehabilitation Agent | Explore AI-assisted rehabilitation while protecting autonomy and dignity | consent boundaries · human escalation · opportunity discovery · privacy |
A privacy-conscious companion and safety platform combining medication support, health monitoring, memories, emotional trends, alerts, and escalation. AI supports caregivers; it does not replace clinical judgment or human relationships.
A voice-first assistive system for blind and visually impaired users. One central design boundary is Smart Silence:
silent → on_demand → balanced → chatty
Danger can override silence. Deciding when AI should not speak is treated as an engineering problem, not a UI preference.
Companion + Coach + Early Warning + Opportunity Engine
A long-term AI companion concept focused on rehabilitation and human potential in constrained environments. It is explicitly not designed as a Judge or Warden.
Status: In Development. Implementation, testing, deployment, and measurement claims will be published only when supporting evidence exists.
Across AgentCraft, I design boundaries around AI, not just capabilities for AI.
- Verified data before generative interpretation when factual correctness matters.
- Deterministic controls around probabilistic models when authority must be constrained.
- Replay and reuse before repeated reasoning when a workflow is already known.
- Observable execution so decisions, tool calls, errors, and costs can be inspected.
- Graceful degradation instead of pretending every model or dependency is always available.
- Explicit human escalation where safety, dignity, or high-impact decisions are involved.
Examples:
- Migration Agent: the LLM understands; verified sources and deterministic rules constrain.
- SaaSForge: AI assembles; the feasibility layer can reject.
- AgentForge: the LLM can reason; deterministic replay can remove repeated reasoning.
- Blind Care Agent: AI can speak; Smart Silence decides whether it should.
- Rehabilitation Agent: AI may support development; it must not judge or punish.
My public GitHub repositories are designed as Engineering Evidence / Technical Case Studies, not as mirrors of private production source.
Each evidence repository is intended to expose the reasoning that matters to technical reviewers:
- architecture and system boundaries;
- engineering decisions and rejected alternatives;
- interfaces and integration contracts;
- testing and verification evidence;
- deployment and controlled-run evidence;
- security and privacy considerations;
- known limitations and technical debt;
- current implementation status.
IMPLEMENTED · TESTED · VERIFIED IN A CONTROLLED RUN · DEPLOYED · MEASURED · ESTIMATED · SCAFFOLDED · PLANNED · NOT YET VALIDATED
Evidence discipline: Estimate ≠ benchmark · Scaffold ≠ completed feature · Roadmap ≠ implementation · Deployment ≠ proof of every capability.
Project Context → Architecture → Engineering Decisions → Evidence → Limitations → Live Demonstration
I intentionally avoid placeholder repository links. Public evidence repositories will be linked as they are completed and reviewed.
I am building the AgentCraft Professional Engineering Portfolio: converting selected systems into rigorous public technical case studies while keeping proprietary production implementation private.
Current work includes architecture documentation, engineering evidence repositories, claim-to-evidence audits, testing records, security notes, live-demo alignment, and cross-project consistency.
Senior AI Engineering · AI Product Engineering · Backend Systems · Agentic Workflows · Full-Stack SaaS · Assistive Technology · Education Technology
I am especially interested in teams building AI systems that must operate under real-world constraints, explicit safety boundaries, measurable engineering tradeoffs, and production accountability.