"My career is a real-world implementation of Reinforcement Learning · no predefined path, just continuous exploration, feedback loops, and optimizing for long-term reward."
I understand what happens beneath the abstraction layers: from transistors to transformers to tokens.
13+ yrs in technology · 7+ yrs in AI/ML
- 🤖 Production Agentic AI: multi-agent orchestration (LangGraph, CrewAI, AG2), durable human-in-the-loop, deterministic control planes, MCP servers & tools (Python + Rust)
- 🔎 RAG at production scale: hybrid + multilingual retrieval, cross-encoder reranking, chunking, anti-hallucination, RAGAS-style evals
- 🎚️ LLM fine-tuning & post-training: domain models via LoRA/QLoRA; GRPO · DPO · offline RL for tool-using agents (applied R&D)
- 💬 Conversational AI & CX: assistants, tool-using agents, chat/voice flows, eval + observability pipelines
- 🛡️ AI-powered offensive security: bug-bounty / VAPT / SAST platforms, Burp + HackerOne MCP, threat modeling (STRIDE / OWASP)
- 📈 Predictive ML & analytics: forecasting, time-series, financial-signal models, ensembles (XGBoost / LightGBM)
- 👁️ Computer Vision & Edge AI: detection & tracking (YOLO + ByteTrack), pose estimation, Jetson / TensorRT optimization
- ⚙️ AI / MLOps: drift monitoring, automated retraining, canary deploys on Kubernetes, distributed training (Ray / DeepSpeed)
- 🏛️ EU-compliant SaaS & architecture: GDPR, EU AI Act, RBAC/RLS, immutable audit trails, reference architectures
- 🧱 Full-stack & backend: Next.js / React, FastAPI / Node, high-concurrency microservices (100k+ users), PostgreSQL / Redis
- 🔄 Automation & data engineering: n8n workflows, ETL (Airflow / Kafka / dbt), scrapers, CRM & billing integration
- ☁️ DevOps & cloud: Docker, Kubernetes, Terraform IaC, CI/CD across AWS / GCP / Azure
🚀 End-to-end ownership: discovery → architecture → build → deploy → evaluation → handoff.
💻 Languages
🧠 LLM & Agents
🎯 RL & LLM Post-Training (applied R&D)
🗄️ Vector DBs & RAG
🔐 AI Security & Offensive
🚀 MLOps & Cloud
⚡ Backend, Data & Frontend
🔄 Automation & Integration
🖥️ AI Dev Tooling
☁️ Proprietary · Claude Opus 4.8 / 4.7 · Sonnet 4.6 · Haiku 4.5 · Fable 5 · OpenAI GPT-5.4 · Codex · Google Gemini 3 Pro / Flash · Embedding 2 · Cohere Command A / R+ · Amazon Nova · Mistral Large / Medium · Azure OpenAI · AI21 Jamba · Perplexity Sonar
🔓 Open-Weight · Meta Llama 4 · 3.3 70B · Alibaba Qwen 3.5 · Qwen3-Coder · DeepSeek V3.2 · R1 · Coder-V2 · Mistral 3 · Mixtral · Codestral · Zhipu GLM-4.6 · Moonshot Kimi K2.6 · Google Gemma 4 · Microsoft Phi-4 · Ai2 OLMo 3 · NVIDIA Nemotron 3 · IBM Granite 3.x
🔎 Embeddings · BGE-M3 · multilingual-e5 · Nomic Embed · gte-Qwen2 · Jina v3 · EmbeddingGemma
🎛️ Fine-Tuning · LoRA · QLoRA · PEFT · Unsloth · TRL (SFT / DPO / GRPO) · quantization (GGUF · AWQ · GPTQ)
🧭 How I use them: model-tier routing · cheap/fast models for recon, routing and extraction; frontier models for hard reasoning and report quality. Self-hosted open-weight (vLLM / Ollama / llama.cpp) for private and air-gapped inference. Provider-agnostic through a LiteLLM gateway, so swapping models is a config change, not a rewrite.
Model-agnostic by default (every major LLM above); deepest in the Claude/Anthropic agentic stack:
| Area | What I use |
|---|---|
| Models | Claude Opus 4.8 / 4.7 · Sonnet 4.6 · Haiku 4.5 · Fable 5 |
| Agentic coding | Claude Code (CLI) · subagents · hooks · slash commands · plugins · CLAUDE.md |
| SDK / API | Claude Agent SDK · Claude API (Messages, tool use, prompt caching, extended thinking, Computer Use) |
| MCP | Model Context Protocol · building MCP servers & clients (Python + Rust) · FastMCP |
| Deploy | Amazon Bedrock · Google Vertex AI · Azure |
Where I go deepest: designing how agents discover tools, coordinate with each other, and stay bounded in production.
🔌 Model Context Protocol (MCP) · building MCP servers & clients in Python and Rust
linkedin-export-mcp(read-only profile reader) ·sentinel-mcp(Rust, in progress: auth, capability-scoped tools, immutable audit log)- Live integrations: Burp Suite + HackerOne MCP inside the Sentinel AI security platform
- Patterns: typed / schema-validated tools, capability scoping, FastMCP, stdio & SSE transports
🤝 Agent-to-Agent (A2A) coordination · how agents delegate, hand off, and negotiate
- Coordinator + specialist topology with typed handoffs, shared state, and approval gates
- Supervisor → worker (Manager-Worker) hierarchies; delegation across agent boundaries
- Built around the emerging interop standards (A2A for agents, MCP for tools)
🧠 Multi-agent systems · best cases
- Sentinel AI · 7 specialist agents, 8-phase pipeline, model-tier routing (Haiku → Sonnet → Opus), 211 tests
- AgentDS · 10-agent Manager-Worker framework with self-correcting reflexion loops + HITL gates
- Autonomous Agentic System · coordinator + specialist, persistent memory, dynamic tool selection, 200+ eval scenarios
- grpo-toolagent · agent tool-selection optimized as a measured objective (GRPO: 100% correctness, 0% safety violations)
🛡️ Production discipline · deterministic control planes, bounded loops, durable human-in-the-loop, eval + observability (tracing by conversation ID), guardrails.
Recent engagement: an air-gapped, multilingual technical-knowledge assistant over 20k-100k PDF datasheets for an industrial manufacturer, fully on-premise, zero data egress, with an ERP read-only sidecar for live stock/lifecycle data.
- Hybrid retrieval: sparse lexical + dense semantic + late-interaction visual (ColPali-style) fused with Reciprocal Rank Fusion (RRF) and a reranker threshold
- Vision-grounded reading on top-k candidate pages: multimodal RAG over scanned + native PDFs; multilingual input (EN · ZH · DE) → chat output (IT · EN)
- Citation-or-refuse, enforced architecturally: multi-agent orchestration refuses any value it cannot verify against an indexed source (anti-hallucination by design)
- Live structured data: read-only ERP sidecar (Indirect Static Read pattern) with routine-driven cache invalidation
- Evaluation goldset (200 questions) as the acceptance gate: ≥98% citation validity · 0% unsupported-answer rate · ≥95% refusal correctness
- On-prem serving: local text + vision LLMs (multi-model, single GPU); citation drawer, freshness badges, AD/SSO
- Compliance baked in: EU AI Act (Limited-Risk, Reg. 2024/1689), NIS2 control map, GDPR Art. 30
Also: hybrid (semantic + BM25) retrieval, cross-encoder reranking, parent-child chunking, query routing, contextual compression, and RAGAS-style eval (context precision/recall, faithfulness) across healthcare, enterprise-document, and conversational-RAG builds.
I design AI systems for global regulatory compliance from day one, not as a retrofit. Every system ships with audit trails, scope enforcement, and compliance documentation baked into the architecture.
- 🇪🇺 EU AI Act: Article 12 immutable event logging, risk classification, GDPR Art. 30/32 governance
- 🇺🇸 NIST AI RMF / CSF 2.0: identity-centric design, least-privilege IAM for non-human agents
- 🌍 ISO/IEC 42001 & 27001: AI lifecycle governance, ISMS integration, continuous evaluation
Applied in: Sentinel AI (NIST/GDPR/ISO 27001) · GrantRadar (GDPR + EU AI Act) · Claude Code Workspace (EU-sovereign infra, Scaleway France).
| 🔵 Data Scientist Insight · Causality |
🟢 ML Engineer Performance · Production |
🔴 AI Engineer Agents · Unstructured Data |
|
|
|
| 🟣 RL Engineer Sequential decisions · applied R&D |
🟠 AI Solutions Architect Systems · Delivery |
⚙️ System Design Scale · Reliability |
|
|
|
🛡️ sentinel-ai-offensive · AI-Powered Offensive Security Platform
7-agent harness for bug bounty / VAPT / SAST with model-tier routing (Haiku → Sonnet → Opus), 25+ integrated tools, 8-phase pipeline, 20 web2 + 10 web3 vuln classes, Burp + HackerOne MCP, NIST/GDPR/ISO 27001 compliance, 211 tests. Delivered to client, then open-sourced.
Python · Multi-Agent · semgrep · nuclei · Pinecone · FastAPI · Docker
🤖 AgentDS · Autonomous Multi-Agent Data Science Framework
10-agent orchestration (LangGraph + Pydantic AI) automating the full ML lifecycle: data cleaning → feature engineering → AutoML → API generation → deploy. Self-correcting reflexion loops, HITL gates, MLflow tracking.
Python · LangGraph · Pydantic AI · LiteLLM · Optuna · FastAPI · K8s
🧠 maitri · Safety-Verified Maternal Referral Co-Pilot
Built on Gemma with a verifier rejection-recovery loop for safe clinical referral guidance.
Python · LLM · Safety · Verifier-loop
📧 InBoxD · AI Email Classification & Auto-Reply
n8n + LLM two-stage pipeline: monitors inbox, classifies vendor/customer inquiries, queries databases, drafts professional replies. Qdrant vector search.
n8n · LLM · Qdrant · Automation
🎯 SportsAnalytics-CV · Real-Time Sports Video Analysis
Multi-object tracking (YOLO v8/v11 + ByteTrack), team classification (K-Means), speed/distance metrics, possession analytics. FastAPI + Streamlit.
PyTorch · YOLO · OpenCV · FastAPI · Docker
🎛️ grpo-toolagent · GRPO from scratch for agent tool-selection (applied RL R&D)
From-scratch GRPO (no RL framework) post-training a tool-using agent on a composite reward (correctness + safety, minus tool-cost & latency). Runs on CPU; reproducible multi-seed results: 100% tool-selection correctness, 0% safety violations vs random/untrained baselines. README maps it to the LLM-scale path (TRL GRPOTrainer + vLLM).
Python · GRPO · RL post-training · eval harness · tool selection
🦀 Coming soon: a production-grade MCP server in Rust · auth, capability-scoped tools, immutable audit log, OpenTelemetry.
Most of my work ships in clients' private repositories under NDA. A representative sample (2022-2026):
| # | Project | Domain | Stack |
|---|---|---|---|
| 1 | GrantRadar · EU-compliant B2B SaaS (21-section architecture, 8-role RBAC + RLS, bilingual, GDPR Art. 30/32, 10k+ concurrent users) | EU SaaS 🇪🇺 | Next.js 15 (App Router) · React · Supabase/Postgres + RLS · Drizzle · Gemini 2.5 · n8n · Stripe · Vercel · Playwright |
| 2 | Claude Code Workspace · managed AI PaaS for EU SMEs (6-layer architecture, per-tenant container isolation, EU-sovereign infra) | AI PaaS 🇪🇺 | Claude Code · LiteLLM gateway · Scaleway Kapsule (K8s) · ArgoCD · HashiCorp Vault · NetBird/WireGuard · Docker |
| 3 | AI System Architecture & Multi-Agent Guide · local MoE-manager + cloud-specialist orchestration, benchmark-driven | Consulting 🇬🇧 | Ollama · vLLM · AG2/AutoGen · llama.cpp · Qwen3-Coder · GLM-4.6 · SWE-bench eval |
| 4 | Conversational RAG + Fine-Tuned Domain LLM · multi-turn memory, SQL-to-text + vector, query routing | NDA | PyTorch · LangGraph · pgvector · cross-encoder rerank · vLLM · FastAPI · Docker |
| 5 | Supply-Chain Demand Forecasting + RAG Assistant · GBM+LSTM ensemble (-18% MAE), 15K-page SOP RAG | Logistics 🇮🇳 | XGBoost · LightGBM · PyTorch (LSTM) · LangChain · Terraform · AWS ECS · Airflow · MLflow |
| 6 | Autonomous Agentic AI System · coordinator+specialist topology, HITL gates, 200+ eval scenarios | Enterprise 🇮🇳 | LangGraph · CrewAI · Pydantic AI · MCP · Claude/GPT · Qdrant · FastAPI · AWS |
| 7 | Enterprise Document Intelligence RAG · hybrid retrieval, cross-encoder reranking, 8K+ docs, sub-3s p95 | Enterprise NDA | LlamaIndex · hybrid BM25+dense · Qdrant/Chroma · contextual compression · FastAPI · AWS |
| 8 | Financial Trading Prediction Platform · ensemble + 12 indicators, drift monitoring, canary deploys | Fintech 🇺🇸 | XGBoost · LightGBM · AWS SageMaker · Evidently · DVC · W&B · Kubernetes (canary) |
| 9 | Healthcare RAG + LLM Fine-Tuning · hybrid search, anti-hallucination, +23% accuracy (LoRA/QLoRA) | Healthcare 🇺🇸 | PyTorch · PEFT (LoRA/QLoRA) · LangChain · BM25+vector · Pinecone · vLLM · GCP Vertex AI |
| 10 | Psychometric Scale Validation · EFA/CFA/SEM, 100+ items, publication-ready report | Academic 🇮🇳 | R (lavaan, psych) · SPSS · Python · SEM/CFA/EFA |
| 11 | Real-Time Pose Estimation · 60+ FPS, 94% mAP for sports analytics | Computer Vision | PyTorch · YOLOv8 · ByteTrack · TensorRT · OpenCV |
| 12 | Edge AI Deployment · NVIDIA Jetson, 4× compression, -85% cloud cost | Edge AI | TensorRT · ONNX Runtime · INT8 quantization · Jetson · Triton |
| 13 | Enterprise GenAI RAG · 10K+ pages, query resolution 15 min → <12 s, streaming | Enterprise | LangChain · Pinecone · FastAPI · Redis · SSE streaming |
| 14 | Predictive Analytics & BI · forecasting/classification, ETL on millions of records, exec dashboards | Healthcare/Logistics | scikit-learn · Airflow · dbt · Snowflake · Power BI · Tableau |
| 15 | Backend Microservices Platform · 100k+ concurrent users, RBAC, Redis (-70% DB load), PostGIS real-time matching (<50ms) | Gov / Delivery 🇮🇳 | FastAPI · Kubernetes (HPA) · PostgreSQL · PostGIS · Redis · Kafka |
| 16 | Network & Security Infrastructure · firewalls, segmentation, topologies, hardening for SMB/enterprise | Networking 🇮🇳 | Cisco · CCNA · firewalls · VLANs · VPN · network hardening |
- AWS · Generative AI and AI Agents with Amazon Bedrock
- IBM · RAG and Agentic AI Professional Certificate
- IBM · Building AI Agents and Agentic Workflows Specialization
- Google · Advanced Data Analytics Professional Certificate
- Google Cloud · Advanced Machine Learning on Google Cloud
"In RL, there are no mistakes, only exploration. Let's build something that ships."