╔══════════════════════════════════════════════════════════════════╗
║ SANTHOSH · AI/ML Engineering Runtime · ACTIVE ║
╚══════════════════════════════════════════════════════════════════╝
$ whoami┌─────────────────────────────────────────────────────────────────────┐
│ IDENTITY BLOCK │
│ ───────────────────────────────────────────────────────────────── │
│ Role : AI / ML Engineer (2nd Year CSE — Karpagam Institute) │
│ Focus : Agentic systems · RAG pipelines · Production MLOps │
│ Output : Deployed systems — not prototypes │
│ Mantra : "Make it work. Make it right. Make it fast." │
└─────────────────────────────────────────────────────────────────────┘
A self-operating system for turning ambiguous problems into deployed AI infrastructure.
╔════════════════════════════════════════════════════════════════════════╗
║ SANTHOSH ENGINEER OS · RUNTIME SPEC ║
╠══════════════╦═════════════════════════════════════════════════════════╣
║ INPUT ║ Problem statements · Research papers · Broken systems ║
╠══════════════╬═════════════════════════════════════════════════════════╣
║ PROCESSOR ║ Agentic reasoning (LangGraph) ║
║ ║ System design thinking → microservice decomposition ║
║ ║ Architecture-first, code-second approach ║
╠══════════════╬═════════════════════════════════════════════════════════╣
║ MEMORY ║ Qdrant / ChromaDB vector stores ║
║ ║ BM25 sparse retrieval + dense embedding fusion ║
║ ║ Past project patterns as reusable mental models ║
╠══════════════╬═════════════════════════════════════════════════════════╣
║ OUTPUT ║ Containerized · CI/CD-wired · Security-hardened ║
║ ║ JSON-logged · Rate-limited · Production-ready ║
╠══════════════╬═════════════════════════════════════════════════════════╣
║ OS TRAITS ║ ✓ Docker from commit one ║
║ ║ ✓ GitHub Actions in every repo ║
║ ║ ✓ Agents / API / ML as separate bounded services ║
║ ║ ✓ Security baked in — never bolted on ║
╚══════════════╩═════════════════════════════════════════════════════════╝
LOADING MODULE REGISTRY... ████████████████████████ 100%
[MODULE 01] AI CORE ENGINE
| Component | Tech | Status |
|---|---|---|
| Agent Orchestration | LangGraph · LangChain | 🟢 ACTIVE |
| Vision & Detection | PyTorch · OpenCV · Transformers | 🟢 ACTIVE |
| STT · Multi-language | Whisper · Sarvam AI Saaras v3 | 🟢 ACTIVE |
| TTS Synthesis | edge-tts · Neural Voices | 🟢 ACTIVE |
[MODULE 02] DATA & RETRIEVAL LAYER
| Component | Tech | Status |
|---|---|---|
| Vector Store | Qdrant · ChromaDB · FAISS | 🟢 ACTIVE |
| Sparse Retrieval | BM25 · Reciprocal Rank Fusion | 🟢 ACTIVE |
| Document Ingestion | MarkItDown · PDF · DOCX · PPTX | 🟢 ACTIVE |
| Embedding Models | MiniLM-L6 · HuggingFace Hub | 🟢 ACTIVE |
[MODULE 03] BACKEND SYSTEMS
| Component | Tech | Status |
|---|---|---|
| API Framework | FastAPI | 🟢 ACTIVE |
| Primary Language | Python | 🟢 ACTIVE |
| Relational Store | PostgreSQL | 🟢 ACTIVE |
| Object Storage | MinIO | 🟢 ACTIVE |
[MODULE 04] INFRASTRUCTURE & DEVOPS
| Component | Tech | Status |
|---|---|---|
| Containerization | Docker · Docker Compose | 🟢 ACTIVE |
| CI / CD Pipelines | GitHub Actions | 🟢 ACTIVE |
| Security Scanning | Bandit · detect-secrets · ruff | 🟢 ACTIVE |
| Log Monitoring | Dozzle · JSON stdout | 🟢 ACTIVE |
[MODULE 05] FRONTEND & UX
| Component | Tech | Status |
|---|---|---|
| Web Framework | Next.js · TypeScript | 🟢 ACTIVE |
| Rapid Prototyping | Gradio | 🟢 ACTIVE |
| Static Generation | Node.js · GitHub Pages | 🟢 ACTIVE |
| Runtime | Linux | 🟢 ACTIVE |
SCANNING PROJECT REGISTRY... 4 SYSTEMS FOUND
Problem: Deepfake media is proliferating faster than detection tooling can scale. Existing solutions are monolithic and brittle — one model, one failure point, no auditability.
Solution: Forensic deepfake detection architected as a production microservices mesh.
flowchart LR
U[Upload] --> API[API Gateway]
API --> AG[Agent Orchestrator]
AG --> ML[ML Service]
ML --> RF[RetinaFace]
ML --> XC[Xception CNN]
ML --> TR[Transformers]
AG --> CH[(ChromaDB)]
AG --> PG[(PostgreSQL)]
PG --> MI[(MinIO Object Store)]
MI --> EX[Auto-Expiry 30d]
style U fill:#5B8DEF,stroke:#333,color:#fff
style API fill:#1a0d0f,stroke:#5B8DEF,color:#fff
style AG fill:#1a0d0f,stroke:#5B8DEF,color:#fff
style ML fill:#1a0d0f,stroke:#5B8DEF,color:#fff
stateDiagram-v2
[*] --> pending
pending --> temp_stored
temp_stored --> processing
processing --> processed
processed --> deleted
deleted --> [*]
Stack: Python · FastAPI · LangGraph · Docker Compose
PostgreSQL · ChromaDB · MinIO · RetinaFace · Xception
CI/CD: GitHub Actions · network audit · logging validation
Architecture: Multi-Dockerfile services (
Dockerfile.api/Dockerfile.agents/Dockerfile.ml) sharing a singledocker-compose.yml. JSON-structured logging across every module with live Dozzle viewer.
Design decision:
FLAG_FOR_REVIEWis treated as an honest middle state, not a failure mode — multi-signal weighted aggregation surfaces uncertainty instead of forcing a binary call.
Problem: Enterprise manuals exist in a graveyard of PDFs and PPTX files. Users either search manually or get generic LLM hallucinations. No voice support for Indic languages. No agentic troubleshooting.
Solution: Production RAG system with hybrid search, voice layer, and LangGraph-powered diagnostic engine.
flowchart TD
A[PDF · DOCX · PPTX · XLSX · TXT] --> B[MarkItDown Parser]
B --> C[Chunker + Overlap Windows]
C --> D[MiniLM-L6 Embeddings]
D --> E[(Qdrant Vector Store)]
flowchart LR
Q[Query] --> M1[Exact Match]
M1 --> M2[Family Match]
M2 --> M3[Global Match]
M1 --> DV[Dense Vector]
M2 --> BM[BM25 Sparse]
M3 --> RRF[RRF Fusion]
DV --> RRF
BM --> RRF
RRF --> LLM[LLM]
flowchart LR
subgraph eng["English Path"]
EA[English Audio] --> W[Whisper STT] --> R1[Response] --> T1[edge-tts]
end
subgraph indic["Indic Path"]
IA[Indic Audio] --> S[Sarvam Saaras v3] --> R2[Response] --> T2[edge-tts]
end
SECURITY STACK:
slowapi rate limiting · MIME-type guards · regex prompt injection
isolated RAG prompts · 24-test suite (unit + integration + security)
Stack: Python · FastAPI · LangGraph · Qdrant · BM25 · MiniLM-L6
Whisper · Sarvam AI · edge-tts · Next.js · TypeScript · Docker
CI/CD: GitHub Actions · ruff · black · bandit · detect-secrets
Architecture: Full-stack monorepo (
backend/+frontend/). Prompt injection filtered at both the regex layer and the LLM system prompt level. Admin upload panel in Next.js.
In progress: Phase 3 hardening — multi-provider LLM key rotation (Gemini → Groq → SambaNova → HF → OpenAI), VLM-based image captioning, parent-child chunking, BGE reranker.
Problem: E-commerce discovery still relies on keyword search. Users describe what they want in natural language — the product catalog doesn't speak that.
Solution: LLM-powered natural language product discovery deployed publicly on Hugging Face Spaces.
flowchart TD
U[User Query - Natural Language] --> G[Gradio UI Layer - app.py]
G --> IL[Inference Logic]
IL --> API[LLM API]
IL --> DL[Data Layer]
DL --> VDB[FAISS / ChromaDB ready]
G --> DK[Docker Container]
DK --> HF[HuggingFace Spaces]
Stack: Python · Gradio · Transformers / LLM APIs · Docker · HF Spaces
CI/CD: GitHub Actions
Architecture: Prompt-orchestration-first design, modular for async inference and vector DB integration with zero structural changes.
Problem: Markdown documentation has no navigable web presence without a full CMS or framework overhead.
Solution: Zero-config static site generator with incremental builds and auto-deploy on push.
flowchart TD
A[Markdown Source] --> B[Node.js Build]
B --> C[HTML + Dark Mode + Search]
C --> D[GitHub Actions]
D --> E[GitHub Pages]
B --> F[Incremental - only changed files regenerated]
Stack: JavaScript · Node.js · GitHub Actions · GitHub Pages
╔══════════════════════════════════════════════════════════════╗
║ SYSTEM METRICS · GITHUB TELEMETRY ║
╚══════════════════════════════════════════════════════════════╝
| Metric | Value |
|---|---|
| Deployed production repos | 4 |
| Independent microservices shipped | 8+ (API · Agents · ML across 2 systems) |
| CI/CD pipelines maintained | 100% of repos (lint · security scan · deploy) |
| Containerized services | 100% (Docker/Compose on every project) |
| Languages in active rotation | Python · TypeScript · JavaScript |
| Voice/ASR languages supported | English · Tamil · Hindi |
Note: the stats/streak/activity widgets above render live from GitHub's public API at view time — they'll always reflect real, current numbers rather than a static snapshot.
Every system in this portfolio ships with the same non-negotiables.
╔═══════════════════════════════════════════════════════════════╗
║ ENGINEERING INVARIANTS ║
╠═══════════════════════════════════════════════════════════════╣
║ [✓] Containerized from commit one — Docker / Compose ║
║ [✓] CI/CD in every repo — lint · security · deploy ║
║ [✓] JSON structured logging — stdout, never print() ║
║ [✓] Security-first APIs — rate limit · MIME guard · scan ║
║ [✓] Separation of concerns — agents ≠ API ≠ ML ║
╚═══════════════════════════════════════════════════════════════╝
// Open to roles in AI engineering, MLOps, and agentic systems.
// I build production systems, not demos. Let's build.
╔══════════════════════════════════════════════════════════════════╗
║ SANTHOSH.OS · SYSTEM HALT · All services nominal. ✓ ║
╚══════════════════════════════════════════════════════════════════╝