Engineering Bounded Autonomy & Production AI Systems.
I’m Muhammad Oliya Mazhar — an 18-year-old AI Systems Architect and independent researcher at COMSATS University Islamabad (BS Artificial Intelligence, 5th Semester). I design deterministic governance engines for multi-agent workflows (BlackSmith), automated financial extraction pipelines (Due), and live 24/7 autonomous client operations (Aivenchy).
Continuous Autonomy Boundary Vector
Real-time state preview of BlackSmith’s floor-clamped governance operator eff(a) = max(B_a, D_a) across 4 enterprise sub-agent action classes. Adjust sliders below in Section 01 to recompute live traces.
BlackSmith: Meta-Orchestration & Bounded Autonomy
BlackSmith is a meta-orchestration framework for dynamic AI sub-agent generation with configurable autonomy boundaries, providing structured governance for multi-agent workflows operating in enterprise environments. Positioned as the first open-source, domain-general framework producing quantitative empirical findings on how autonomy boundary configuration affects agent reliability — addressing a critical research gap left unaddressed by commercial platforms such as Vertex AI Agent Builder, Salesforce Agentforce, and Azure AI Foundry.
Continuous boundary configuration vector governing Autonomous, Assisted, and Halted decision states per action class, layered with a Decision Checkpoint Set enforcing hard-stop conditions for irreversible operations independent of B.
Intercepts every agent decision node before execution, evaluating actions against global halt conditions, computing the effective boundary via a floor-clamped max(B, D) operation, and routing to tool execution or a structured Human-in-the-Loop checkpoint.
Executes structured scenario suites across Normal, Edge Case, Adversarial, and Failure Injection categories to stress-test generated sub-agents under varying governance constraints without live production risk.
Iteratively refines autonomy configurations via numerical gradient estimation, improving Task Completion Rate, Decision Consistency Score, and Failure Mode Containment Rate while respecting D-value safety floors.
Why BlackSmith Exists vs. Commercial Suites
Commercial orchestration platforms offer static guardrails but lack a continuous mathematical formulation of autonomy boundaries or reproducible datasets measuring the autonomy-reliability tradeoff.
| CAPABILITY | BLACKSMITH | VERTEX / AGENTFORCE / FOUNDRY |
|---|---|---|
| Continuous Vector B ∈ [0,1] | Yes (Per Action Class) | Binary / Static Rules |
| Floor-Clamped max(B, D) | Native Node Interceptor | Not Formulated |
| Gradient Boundary Optimiser | Numerical Gradient Est. | Manual Prompt Tuning |
| Execution Trace Dataset | Reproducible SQLite Log | Proprietary Telemetry |
| Open-Source Empirical Benchmark | 100% Open-Source | Closed Commercial SaaS |
Sandboxed Evaluation & Boundary Optimiser Simulator
Production Engineering: Due & Aivenchy
Automated OAuth 2.0 Gmail Invoice Extraction & Executive Spend Telemetry
Connects to Gmail through OAuth 2.0 to automatically pull PDF invoices out of incoming mail, alongside a direct drag-and-drop PDF upload pipeline. Uses OpenAI’s structured output parsing (gpt-4o-mini) to extract vendor, invoice number, due date, amount, currency, and payment notes without manual entry — backed by a FastAPI + SQLite REST API for full create, edit, search, filter, and status updates.
| INVOICE # | VENDOR | DUE DATE | AMOUNT | STATUS | REST API ACTION |
|---|
Live view of the schema-validated JSON payload extracted from incoming PDF attachments and committed to SQLite via FastAPI endpoints:
24/7 Autonomous WhatsApp Front-Line, Missed-Call Recovery & Human Handoff
Live and running in production for a real client at aivenchy.lavenchy.com, handling actual customer conversations around the clock. Detects missed calls and messages the caller on WhatsApp within seconds before the lead goes cold, answers common questions, books appointments autonomously after checking the existing schedule so nothing double-books, and hands off to a human agent whenever a request is outside its scope or the customer asks for one.
Missed Call Instant Recovery (< 2.4s)
Primary & Secondary Engineering Stack
COMSATS University Islamabad — BSAI (5th Semester)
Bachelor of Science — Artificial Intelligence
Currently enrolled in 5th Semester at COMSATS University Islamabad (Islamabad Campus), combining rigorous mathematical foundations in Linear Algebra, Probability & Statistics, and Algorithm Design with independent empirical research in multi-agent governance.
BlackSmith was conceived, formally proposed, and is being built entirely independently — from initial research proposal through mathematical boundary formulation and architecture design to full Python/LangGraph implementation — demonstrating end-to-end research ownership at 18 years old.
Click any of the 7 core BSAI coursework modules below to inspect how university theory is applied directly in BlackSmith, Due, and Aivenchy:
Initiate a Technical Conversation.
Available for AI Systems Engineering roles, Agentic Governance research collaborations, and production LLM/automation architecture mandates. Based in Islamabad, Pakistan and collaborating globally.