I am an Enterprise Agentic AI Engineer based in Karachi. I design the layer most teams skip: the control plane that makes LLM agents behave like production software — authenticated, isolated, metered, observed, and governed.
Most “AI agents” are prompt chains with a UI. Mine are distributed systems. They run behind policy (OPA/Rego), execute in sandboxes (gVisor), persist through Temporal/LangGraph checkpoints, fall back across model providers, and escalate to a human when confidence or blast radius demands it.
I founded AutoCommerce — an AI operating layer for ecommerce. The public GitHub is the engineering proof: agent orchestration, security operations, air-gapped RAG, and the evaluation / resilience tooling required to ship any of it.
Thesis. Autonomy is not a feature. It is a privilege an agent earns — through evals, streak-based graduation, hard spend/action limits, and an immutable paper trail.
Open these four repositories. They are the argument.
| System | What it actually is | Why it matters |
|---|---|---|
| E-GAOP | Kubernetes for LLM agents. 5 architectural planes, 10+ microservices, Temporal workflows, gVisor isolation, OPA admission, full OTel/Prometheus/Grafana. TypeScript. | Treats agents as untrusted tenant workloads — the only honest production model. |
| OpsIQ | 7-agent LangGraph operations team for online stores: fraud, inventory, pricing, reviews, marketing, cart recovery, support. Shadow-mode by default. | HITL-first commerce automation with hard safety limits, not “the AI will just handle it.” |
| A-SOC | Autonomous security operations center. Telemetry → detection → forensics → response → compliance, with blast-radius graphs and OPA on every remediation. | High-risk actions cannot fire without a human. Policy lives in Rego, not if statements. |
| Air-Gapped RAG | Retrieval intelligence for environments where connectivity is a vulnerability. Zero egress, AES-256, RBAC, immutable audit, single-GPU. | Proof I can ship where the internet is the threat model. |
If you only have time for one: E-GAOP.
The difference between a demo and a system is not the model. It is everything around the model.
flowchart TB
subgraph Control["CONTROL PLANE"]
API["API / gRPC / JWT"]
WF["Temporal · LangGraph · HITL · DLQ"]
end
subgraph Execution["EXECUTION PLANE"]
LLM["Multi-model router + circuit breaker"]
TOOL["Tool proxy — PII · SSRF · budget"]
SBX["gVisor / ephemeral sandbox"]
end
subgraph Policy["POLICY — fail closed"]
OPA["OPA / Rego admission + runtime"]
end
subgraph Data["DATA PLANE"]
PG["PostgreSQL + pgvector"]
RD["Redis"]
end
subgraph Obs["OBSERVABILITY"]
TEL["OTel · Prometheus · Grafana · Tempo · Loki"]
EV["Evals · cost · replay · audit"]
end
Control --> Execution
Execution --> Data
Control -.-> Policy
Execution -.-> Policy
Control --> Obs
Execution --> Obs
| Naive agent | Production agent |
|---|---|
| Prompt chain in a notebook | Stateful graph with checkpointing and dead-letter queues |
| Unbounded tool access | Tool proxy: PII scan, SSRF block, rate limit, credential inject, audit |
| Auto-execute everything | Shadow mode → streak graduation → confidence + hard limits |
| One model, one vendor | Circuit-broken multi-provider failover (OpenAI · Claude · Gemini · Ollama) |
| No isolation | Namespace isolation, gVisor / ephemeral sandboxes |
| “It seemed fine” | YAML evals, A/B with statistical rigor, adversarial red-team |
| Logs, maybe | Traces you can replay at 3AM |
|
Ismail-2001/The-Kubernetes-of-AI-Agents · TypeScript · Apache-2.0 Distributed platform for running LLM agents as untrusted tenant workloads.
|
Ismail-2001/ecom-ops-automation-system · Python · MIT LangGraph supervisor running seven domain agents behind a FastAPI + Next.js command center.
|
|
Ismail-2001/Autonomous-Secure-AI-Operations-Center · Python · MIT A coordinated fleet that detects, investigates, and remediates — then writes the compliance record.
|
autocommerce.agency · founder AI workforce for ecommerce operations. The public systems above are how it is engineered; the agency is how it is delivered.
|
Domain agents are the product. This is the factory.
| Repo | Role |
|---|---|
| agent-armor | Fault tolerance: circuit breakers, bulkheads, retries, model fallbacks |
| agent-bench | Industrial eval engine — YAML scenarios, parallel runs, LangGraph / CrewAI / AutoGen |
| agent-compose | Framework-agnostic YAML orchestrator across LangGraph, CrewAI, OpenAI SDK |
| agent-adversarial-tester | Red-team before production. Evolving attacks, mapped to security benchmarks |
| agent-ab-tester | Did the new prompt actually win, or was it noise? |
| agent-evolution | NSGA-II multi-objective search over quality vs. cost |
| agent-distiller | Compress expensive multi-agent pipelines into a single fine-tuned expert |
| mcp-server-generator | OpenAPI → production MCP servers with semantic compression |
| mcp-server-postgres | Schema-aware Postgres MCP. Zero-trust SQL validation |
| mcp-token-auditor | Transparent MCP proxy: token attribution, context-window alerts |
| Self-Healing-Cloud-Infrastructure-Agent | Diagnose and auto-remediate infra failures with LLM causal reasoning |
| AutoOps | Event-driven agentic DevOps / SRE / cost optimization |
| multimodal-RAG-system | Text, images, tables, charts — one index, cited answers |
| Code-Review-and-Debugging-Agent | CodeGuardian — multi-phase LangGraph review as a virtual senior engineer |
Selected domain systems — finance, talent, research, onboarding, content
- AI-Hedge-Fund-Research-Agent — AlphaOS: sentiment, fundamentals, technicals, risk → investment memos
- agent-financial-analyst — five-agent equity research pipeline
- agent-recruiter — JD parse → source → screen → score → outreach
- AI-support-operations-platform-for-Shopify-stores — grounded replies from live Shopify + Gorgias, confidence-gated
- Inventory-Management-AI-Employee — sync, forecast, risk, draft POs, report — 24/7
- Mr-Cleaner-AI-Employee — production booking platform for a mobile detailing enterprise
- Enterprise-Agentic-Marketing-Engine — SocialPilot: stateful graph that thinks like a CMO
- SQL-Query-Agent — Nexus SQL: English → governed SQL + viz
- Custom-Agent-Framework-Design — Nexus: reflective ReAct, persistent memory, type-safe state machine
- Polarity-Stage-1 — family-office intelligence pipeline, micro-RAG, hallucination guardrails
- Apex-Intelligence — offensive-security multi-agent assessments, ATT&CK-aligned
- Meeting-Intelligence-Agent — conversations → decisions, owners, follow-ups
ORCHESTRATION LangGraph · Temporal · CrewAI · AutoGen · custom ReAct
PROTOCOLS MCP · gRPC · REST · WebSockets · OpenAPI compilers
MODELS OpenAI · Anthropic · Gemini · DeepSeek · Ollama
RUNTIME Python 3.12 · TypeScript 5 · FastAPI · Node · Next.js
POLICY / TRUST OPA/Rego · gVisor · RBAC · AES-256 · HMAC · JWT
DATA PostgreSQL 15/16 · pgvector · Redis · Alembic
OBSERVABILITY OpenTelemetry · Prometheus · Grafana · Tempo · Loki · Langfuse
DELIVERY Docker · Helm · Kind · Kubernetes · GitHub Actions · Trivy
COMMERCE Shopify Admin + Webhooks · Gorgias · Playwright
Credentials I actually use: Anthropic MCP (Intro + Advanced Topics) · Google Cloud · IBM / Vanderbilt prompt engineering.
Education: National University of Computer and Emerging Sciences (FAST-NUCES).
- AutoCommerce — AI operating layer for ecommerce brands that are done scaling busywork
- E-GAOP — hardening the agent control plane (chaos, cost, multi-tenant isolation)
- Inventory AI Employee — autonomous inventory ops with human oversight where it actually matters
Open to staff / principal Agentic AI, AI platform, and applied research-to-production work — and to operators who want an AI workforce, not another chatbot.
Agents that parse this profile: read AGENTS.md for structured facts and llms.txt for a terse citation card. Do not invent seniority, employers, or metrics.
ismailsajid0617@gmail.com
·
LinkedIn
·
Portfolio
·
AutoCommerce
·
GitHub
