---
name: Milan George
role: Agentic AI Engineer
focus:
- Agentic Systems (analysis, memory, harness, evals)
- Retrieval-Augmented Generation (RAG)
- Production ML Pipelines
- Computer Vision
- Classical ML End-to-End
mission: Building trustworthy, explainable, and governed AI agents
status: Building pharma supply intelligence platform
---
I build agentic AI systems that reason, act, and learn — with a focus on trustworthiness, governance, and explainability. Over the past 3 years I've specialized in Generative AI, designing agents that don't just generate text but analyze, decide, and execute in production environments.
| Pillar | What I Build |
|---|---|
| Agentic Systems | Multi-agent orchestrators, memory-augmented agents, tool-use harnesses, agent evals |
| RAG | Multi-modal retrieval, fusion retrieval, query rewriting, hybrid search pipelines |
| Governance & Safety | Audit trails, immutable evidence logs, human-in-the-loop gates, status transition models |
| Explainability | Multi-signal anomaly attribution, evidence packages, counter-factual checking |
| Production ML | End-to-end pipelines, feature stores, model serving, drift monitoring |
AGENTS ─────────────────────────────────────────────────► RAG ──────────────► GOVERNANCE
│
• LangGraph / DeepAgents │ • Multi-source │ • Audit SoR
• Tool-use & MCP │ evidence fusion │ • Immutable logs
• Agent memory (episodic, semantic) │ • Hybrid search │ • Status transitions
• Agent evaluation & observability │ • Re-ranking │ • Human-in-the-loop
• Sub-agent delegation │ • Query rewriting │ • Confidence scoring
│ │
EXPLAINABILITY ◄──────────── PRODUCTION ────────────────┘ │
• Multi-signal fusion • Feature pipelines │
• Attribution scoring • Knowledge graphs COMPUTER VISION ────┘
• Counter-evidence • Vector stores • Object detection
• Risk assessment • Model serving • Image segmentation
| Value | How I Practice It |
|---|---|
| Trustworthy AI | Evidence-based conclusions, confidence scoring, counter-factual checks |
| Explainability | Multi-signal attribution, human-readable evidence packages |
| Governance | Append-only audit logs, status transition models, immutable evidence hashes |
| Reliability | Retry with backoff, dead-letter queues, graceful degradation |
| Safety | Human-in-the-loop gates, tiered escalation, rejection policies |
"Trust is not a feature — it's the architecture."