I design and ship production AI systems: multi-agent platforms, RAG pipelines, fine-tuned LLMs and vision-language models, deployed on Azure / AWS / GCP, and I build them security-aware.
📄 Indian Patent Filed · 📚 IEEE Xplore Published · 🎓 M.Tech AI (PG) completed, SRM (CGPA 9.6/10) · 🚀 Live AI product + active startup build · 🛡 Cyber-security labs on GitHub · 🛠 9+ yrs safety-critical engineering before AI
📍 Chennai, India · open to Bangalore · Hyderabad · Remote · ⚡ Immediate joiner
| Role | What I bring |
|---|---|
| GenAI / AI Architect | End-to-end design: RAG, LLM routing, evaluation, cloud deployment, IaC, compliance |
| Agentic AI Architect / Engineer | Hierarchical multi-agent systems (Google ADK), tool calling, memory, LLM-as-Judge evals |
| GenAI / LLM Developer | Fine-tuning (LoRA/QLoRA), RAG with Qdrant/FAISS, FastAPI services, production LLM apps |
| ML / Computer Vision Engineer | PyTorch, CLIP, BLIP, ViT, OpenCV; classical ML with XGBoost / scikit-learn |
| AI Security / DevSecOps (AI-focused) | Hands-on labs in DevSecOps on Azure, DFIR, SOC monitoring (Wazuh), web/API testing and AI/LLM security |
| Research Engineer | Patent filing + IEEE Xplore paper with benchmarked results |
A 3-level agent hierarchy that turns a plain-English goal into tested, reviewed, committed code. Google ADK · Gemini 2.5 Flash · Multi-Agent · LLM-as-Judge · Git automation
flowchart TD
G[Engineer's goal in plain English] --> O[L0 Orchestrator<br/>≤7 feature specs]
O --> F1[L1 Feature Lead<br/>≤8 task specs]
O --> F2[L1 Feature Lead]
F1 --> E[Explorer]
F1 --> C[Coder]
F1 --> T[Tester]
F1 --> R[Reviewer]
R --> GC[(git commit per task)]
- Versioned markdown specs as the contract between agents
- Path-sandboxed file tools, YAML-configurable models / tools / retries
- Golden-task evaluation + LLM-as-Judge scoring, live status board with full traces
- Delivered a HIPAA-oriented healthcare QA system (JWT auth, audit logging, PHI/PII protection)
🏥 Anbu Health AI (live in production)
Multilingual (Tamil/English) AI health assistant used by real patients at a village clinic, with voice input, OTP login and medical report/prescription analysis. Delivered as a paid engagement under AI Vision.
FastAPI · React · Groq LLaMA 3.3 70B + Gemini fallback · GPT-4o vision · Qdrant RAG · Supabase · Azure Container Apps · Terraform · GitHub Actions · Prometheus/Grafana
flowchart LR
U[Patient<br/>Tamil voice / text] --> FE[React frontend]
FE --> API[FastAPI on Azure Container Apps]
API --> AUTH[Firebase phone OTP]
API --> RAG[Qdrant medical RAG]
API --> LLM[LLaMA 3.3 70B → Gemini fallback]
API --> VIS[GPT-4o vision<br/>reports & medicine strips]
API --> DB[(Supabase: history, consent, audit)]
API --> TTS[Tamil TTS]
API -.-> MON[Prometheus / Grafana]
DPDP Act 2023 compliance · consent management · PHI redaction · right to erasure · Schedule H1 drug-safety blocking.
Patent filed (App. No. 202641043947) · IEEE Xplore, ICSTSDG 2026 2026 2nd International Conference on Smart Technologies for Sustainable Development Goals, Chennai · 19–21 Aug 2026 · added to IEEE Xplore 06 Oct 2026 · DOI 10.1109/ICSTSDG69220.2026.11708384
Monolithic prompting spends the same compute on every query. Antahkarana is a five-module cognitive architecture for LLMs and VLMs that routes by question type and only triggers verification / self-consistency passes when needed.
flowchart LR
Q[Query] --> M[Manas<br/>6-way question-type router]
M --> C[Chitta<br/>dense retrieval +<br/>entity-overlap scoring]
C --> B[Buddhi<br/>3-pass reasoning<br/>verify ≈8% · self-consistency ≈12%]
B --> A[Ahamkara<br/>structured output]
S[Sakshi<br/>non-interfering logging monitor] -.-> M
S -.-> C
S -.-> B
S -.-> A
| Result (Qwen2.5-7B / BLIP-2 Flan-T5-XL) | Outcome |
|---|---|
| TruthfulQA accuracy vs direct prompting | +59.61% |
| SVAMP exact match | +40.06% |
| FEVER F1 vs CoT | +250.82% |
| Model calls vs uniform Self-Consistency (VQA) | 26% fewer (2,219 vs 3,000) |
| VQA hallucination | 26.6% → 20.9% (79.1% exact match) |
On HotpotQA and MMLU it performs comparably to, or below, Self-Consistency, an intentional precision-vs-overcall calibration documented in the paper. Live demo
Five hands-on security-engineering labs built to demonstrate practical blue-team and security-engineering skills.
| Lab | Focus |
|---|---|
devsecops-azure-lab |
DevSecOps on Azure |
dfir-investigation-lab |
Digital forensics & incident response |
mini-soc-lab |
SOC monitoring |
web-api-security-assessment |
Web / API penetration-testing assessment |
ai-security-lab |
AI / LLM security |
Python · Azure + Terraform · Docker Compose · GitHub Actions CI/CD · Wazuh SIEM/XDR · Ollama (local LLM) · MIT
| Project | Highlights |
|---|---|
| AgentNet Enterprise Support | Triage → Response → LLM-as-Judge self-grading loop with vector memory (LangChain, FAISS, Vertex AI) |
| Antahkarana Medical AI | Multimodal medical assistant on Azure with CI/CD (LangChain, Azure OpenAI, FastAPI) |
| CodeLlama Coding Assistant | QLoRA fine-tune of CodeLlama-7B on AWS SageMaker, ~60% GPU memory reduction · Demo |
| TinyLLaMA & Mini-VLM | LLM and VLM trained from scratch in PyTorch (loss 8.9→0.33 and 17→1.17), ~$10 compute |
| MML Smart Campus Security | Zero-shot CLIP + BLIP VQA + voice biometrics for CCTV monitoring |
| Conscious AI Browser | Chromium built from source (26k build steps) with a LLaMA 3.3 70B summarization sidebar |
| Hospital Readmission Predictor | Two-stage XGBoost pipeline with SMOTETomek balancing and VIF feature selection |
| Retro Pong (Android) | Kotlin + Jetpack Compose Atari-style game: 3 difficulty levels, saved stats, English / Tamil / Hindi; no permissions / no network / no data collected |
Founder & AI/ML Engineer: AI Vision (Udyam-registered MSME) · Jul 2023 – Present Independent GenAI / agentic AI consulting and product delivery: Anbu Health AI (paid client engagement: build, deploy, monthly maintenance), a spec-driven multi-agent platform for a startup, and aivisionlabs.tech.
Electrical Constructional Site Engineer: SR Electrical Works · 2017 – 2023 Electrical Maintenance Engineer: Mod Forge Pvt Ltd · 2014 – 2016 9+ years running safety-critical installations (500 kVA transformers, 400 kW motors). That's why I design AI with guardrails, observability and failure modes in mind.
| GenAI / Agents | Google ADK · LangChain · RAG · Qdrant · FAISS · Prompt & context engineering · Tool calling · Agent memory · LLM-as-Judge evaluation |
| LLM Training | PyTorch · Hugging Face Transformers · PEFT · LoRA · QLoRA · bitsandbytes · TRL |
| Models | Gemini · LLaMA 3.x · Qwen · GPT-4o · CodeLlama · BLIP · CLIP · ViT · Ollama (local) |
| Computer Vision | OpenCV · CLIP · BLIP · Vision Transformers · multimodal VQA |
| Classical ML | scikit-learn · XGBoost · TensorFlow · statsmodels |
| Backend | Python · FastAPI · REST · SQLAlchemy · Firebase Auth · JWT |
| Cloud & MLOps | Azure Container Apps · AWS SageMaker/S3/EC2 · GCP Vertex AI · Docker · Terraform · GitHub Actions · Prometheus · Grafana · wandb |
| Data | PostgreSQL · MySQL · MongoDB · Supabase · pandas · ETL |
| Security & Compliance | DPDP Act · PHI redaction · consent & audit logging · DevSecOps · DFIR · SOC monitoring (Wazuh SIEM/XDR) · web/API security testing · AI/LLM security |
| Other | Kotlin · Jetpack Compose · JavaScript/React · Linux · Bash · Git |
- M.Tech, Artificial Intelligence (postgraduate, completed): SRM Institute of Science and Technology, CGPA 9.6/10 (May 2026)
- IITM Pravartak Certified AI & ML Professional Program: GUVI × IIT Madras Pravartak × NASSCOM
- B.E., Electrical & Electronics Engineering: Thangavelu Engineering College
- AWS Solutions Architect Associate (Udemy course certificate) · Google × Kaggle AI Agents Intensive · DevOps (55.5 hrs) · OpenCV University · PyTorch · TensorFlow
- NPTEL Top 5%: Introduction to IoT (IIT Kharagpur, 50,282 candidates)
- 4× Cyber Security courses (Hackers Exposed, Network Security, Anonymous Browsing, End Point Protection; 54.5 hrs), plus the five labs above
- 📄 Indian Patent filed (Apr 2026) + IEEE Xplore paper (DOI 10.1109/ICSTSDG69220.2026.11708384)
- 🤗 3 public models on Hugging Face
- 🚀 Production AI product serving real patients in rural Tamil Nadu
- 📱 Native Android game shipped (Kotlin + Jetpack Compose)
- 🏆 SRM Hackathon: 6th of 195 · 100% attendance across all M.Tech semesters
- 📚 Google Scholar · ORCID · YouTube @TheArtificialSI (Artificial SI: AI education in Tamil)
If you're hiring for GenAI / Agentic AI architecture, LLM engineering, ML or computer vision, I'd love to connect.
✉️ rajaganaa@gmail.com · 💼 LinkedIn · 🌐 rajaganaa.github.io