| 🎯 Role | AI Engineer · GenAI Builder · Backend Architect |
| 🔍 Focus | RAG Systems · LLMs · Real-time AI Pipelines |
| 🧠 Mindset | Systems thinker. Builder. Solver. |
| 🚀 Superpower | Production-grade AI — not just notebooks |
| ⚙️ Currently | Designing LLM-augmented pipelines |
| 📚 Learning | Multimodal RAG · Agent Frameworks |
| 🤝 Available | Collaborations & Open Source |
One-line summary: Building production-grade GenAI systems using RAG, local LLMs, and scalable backend architectures.
Production-style Retrieval-Augmented Generation pipelines — not toy projects.
flowchart TD
A[🧑 User Query] --> B[🔍 Vector Search\nChromaDB · all-MiniLM-L6-v2]
A --> C[🔑 Keyword Search\nBM25 · rank_bm25]
A --> D[🗂️ Metadata Filter]
B --> E[VECTOR_TOP_K]
C --> F[BM25_TOP_K]
D --> G[META_FILTER]
E --> H[🎯 Hybrid Merger\nFINAL_TOP_K]
F --> H
G --> H
H --> I[🤖 Local LLM via Ollama\nmistral · qwen2.5-coder]
I --> J[📤 Context-Aware Answer]
style A fill:#1e1b4b,color:#fff,stroke:#7C3AED
style H fill:#312e81,color:#fff,stroke:#A78BFA
style I fill:#1e1b4b,color:#fff,stroke:#7C3AED
style J fill:#14b8a6,color:#fff,stroke:#0d9488
Key features built:
- ✅ Follow-up query detection & topic anchoring
- ✅ Session-based context memory via Redis (
chat:session:{user_id}:{session_id}) - ✅ Intelligent routing — Casual vs Technical vs Follow-up queries
- ✅ Multi-stage retrieval pipeline
- ✅ Real-time chat UI with Streamlit
- ✅ Automotive domain QA & Policy/Document QA systems
- 🔜 LLM-based query rewriting (next-level in progress)
Real-world AI-powered safety monitoring system with real-time event pipelines.
flowchart LR
A[📷 Camera Feed] --> B[🤖 ML Detection Engine\nPPE · Zone Breach]
B --> C[POST /api/v1/violations]
C --> D[🔥 Real-Time Pipeline\nBase64 Decode · Snapshot Store]
D --> E[⚡ Socket.IO Broadcast]
E --> F[🖥️ Next.js Dashboard\nLive Alerts · Role-based UI]
F --> G[👤 Admin]
F --> H[👤 Manager]
F --> I[👤 Supervisor]
style A fill:#1e1b4b,color:#fff,stroke:#7C3AED
style B fill:#312e81,color:#fff,stroke:#A78BFA
style D fill:#1e1b4b,color:#fff,stroke:#7C3AED
style F fill:#14b8a6,color:#fff,stroke:#0d9488
Intelligent hiring system. No hardcoded skill lists. Just LLM-powered understanding.
flowchart TD
A[📄 Resume Upload\n150+ resumes] --> B[Phase 1: Parsing\nMetadata Extraction]
B --> C[🧠 Dynamic Skill Extraction\nNo hardcoding — LLM-powered]
C --> D[Phase 2: JD Matching\nFast Inference]
D --> E[🎯 Context-Aware Scoring]
E --> F[📊 LLM-Based Evaluation\nRanked Output]
style A fill:#1e1b4b,color:#fff,stroke:#7C3AED
style C fill:#312e81,color:#fff,stroke:#A78BFA
style F fill:#14b8a6,color:#fff,stroke:#0d9488
Designed for 10–20 concurrent users, offline-first AI, multi-worker queuing.
flowchart LR
A[🖥️ UI] --> B[⚡ FastAPI]
B --> C[📬 Queue]
C --> D[👷 Workers]
D --> E[🔴 Redis\nSession Memory]
D --> F[🧠 RAG Pipeline]
F --> G[🤖 LLM]
E --> F
style A fill:#1e1b4b,color:#fff,stroke:#7C3AED
style C fill:#312e81,color:#fff,stroke:#A78BFA
style F fill:#312e81,color:#fff,stroke:#A78BFA
style G fill:#14b8a6,color:#fff,stroke:#0d9488
💡 Bottleneck-aware: LLM inference = slowest part → Worker scaling strategy applied.
Models I work with: mistral · qwen2.5-coder · all-MiniLM-L6-v2
| Domain | Topics | Status |
|---|---|---|
| RAG Architectures | Naive · Hybrid · Tool-Augmented · Multimodal | ✅ Built + Learning |
| Embeddings | Semantic search · Sentence Transformers | ✅ |
| Chunking Strategies | Sliding window · Semantic chunking | ✅ |
| Retrieval Evaluation | Precision · Recall · MRR challenges | ✅ |
| LLM Limitations | Hallucination control · Context limits | ✅ |
| Session Memory | Redis-based · Per-user isolation | ✅ Built |
| Offline-first AI | Local LLMs · No cloud dependency | 🚀 Rare skill |
| Trait | Details |
|---|---|
| 🏗️ End-to-end builder | Not just models — full systems from UI to DB to LLM |
| 🚀 Offline-first AI | Local LLMs, no cloud dependency — rare in the field |
| ⚡ Real-time pipelines | Socket.IO + async queuing for live AI events |
| 🧩 System design first | Think in bottlenecks, scale, and worker architecture |
| 🔍 Hybrid retrieval | Vector + BM25 — not just naive vector search |
| 🛠️ Production mindset | Debugged CORS, Prisma, Redis, API failures in real deployments |
| 🔜 Upcoming | Details |
|---|---|
| 🧠 LLM Query Rewriting | Rewrite ambiguous queries before retrieval |
| 🖼️ Multimodal RAG | Vision + Text retrieval pipelines |
| 🤖 AI Agent Frameworks | Tool-use, planning, memory-augmented agents |
| 📊 Evaluation Pipelines | RAGAs · TruLens benchmarking |
| 🚀 Performance Dashboards | Latency · Throughput · Retrieval quality metrics |
"I don't just use AI. I build the systems that power it." 🚀