<!DOCTYPE html>
AI / ML ENGINEER · LLM SYSTEMS · AGENTIC AI · RAG
I spent nine years engineering systems where failure wasn't an option. Then two years building real AI products for real clients — anbuclinic.me is live today, billed and running in Tamil Nadu. A filed patent and an IEEE paper are the research proof.
Most AI engineers understand models.
I understand failure — the real kind,
in 400 kW industrial systems where downtime
costs money and safety.
That discipline is in every line of AI code I write.
I am an AI/ML Engineer and the founder of AI Vision (Udyam Reg: UDYAM-TN-02-0483528) — a registered freelance AI practice I've been running for 2+ years alongside my M.Tech. I don't just build AI systems as projects. I build them for paying clients, invoice them, and keep them running in production.
My most recent work is Anbu Health AI — live today at anbuclinic.me. It's a real AI doctor assistant for village clinic patients in Ariyalur District, Tamil Nadu — Tamil voice + text, medicine image analysis (GPT-4o), lab PDF parsing, Groq LLaMA 3.3 70B, Qdrant RAG, Sarvam Tamil TTS, Azure Container Apps, full DPDP Act compliance. Invoice AIV-2026-001 issued to Anbu Clinic in June 2026. This is paid freelance work.
My flagship research project, Antahkarana, is a cognitively-inspired adaptive reasoning framework for LLMs and VLMs — drawing on Vedantic cognitive architecture. It earned a filed Indian patent and a submitted IEEE Conference paper now indexed on Google Scholar and linked to my ORCID researcher identity.
What I bring that most candidates don't: 9 years of safety-critical engineering discipline, 2+ years of real freelance delivery, a live billed product, a patent, and an IEEE paper — all built while completing an M.Tech with a 9.6 CGPA.
A cognitively-inspired, modular conditional reasoning framework for LLMs and VLMs. Inspired by Vedantic cognitive architecture, Antahkarana routes complex queries through specialised stages — perception, discrimination, memory, and integration — to produce coherent, grounded responses at scale.
A real AI doctor assistant for village clinic patients in Ariyalur District, Tamil Nadu — built under AI Vision (my registered freelance company) and billed to Anbu Clinic. Invoice AIV-2026-001 · ₹21,500 · June 2026. Tamil voice + text, medicine image analysis, lab PDF parsing, appointment booking. DPDP Act 2023 compliant. Running in production today.
AI-driven surveillance and threat detection using Vision-Language Models (CLIP, BLIP) and Voice Biometrics. Solves the challenge of monitoring hundreds of CCTV feeds simultaneously.
Autonomous, self-correcting support system using multi-agent orchestration. Agents triage, respond, and self-grade output quality using LLM-as-a-Judge. Vector retrieval memory for contextual recall.
Predicts 30-day hospital readmission risk using AI-driven A1C imputation. Targets the $41B annual cost of preventable US readmissions. Clinical feature engineering + XGBoost pipeline.
Automated ETL pipeline ingesting, cleaning, transforming, and loading YouTube creator data into a structured MySQL warehouse. Analytics-ready output for engagement trend analysis.
Geo-spatial + NLP dashboard mapping India's industrial workforce distribution across sectors. Combines choropleth visualisation with natural language querying for policy planning.
Comprehensive analysis of India's digital payment ecosystem using PhonePe Pulse data. Visualises state-level adoption, transaction volumes, and payment category trends.
Built and trained a full LLM and Vision-Language Model from scratch in a single day. Transformer theory → LLM (loss 8.9→0.33) → VLM (loss 17→1.17) → LoRA fine-tune (loss 1.6→1.01) → published live on HuggingFace. Both models public at RajGana/tinyllama-alpaca-finetuned and RajGana/mini-vlm-scratch.
Compiled Chromium from source (26,000 build steps) on GCP. Custom branded browser with built-in AI sidebar powered by LLaMA 3.3 70B — summarizes and explains any webpage in real-time. Packaged as Linux .deb installer. Same architecture as Brave and Arc browser.
Fine-tuned CodeLlama-7B with QLoRA (4-bit) on CodeAlpaca-20K using AWS SageMaker ml.g4dn.xlarge. Full MLOps pipeline: SageMaker → S3 → HuggingFace Hub → Gradio Space → Groq inference API. Live demo supports code completion, debugging, and explanation in real time.
Introduces a modular conditional reasoning framework for large language models and vision-language models, inspired by Vedantic cognitive architecture (antahkarana). The system routes complex queries through four specialised cognitive stages — manas (perception), buddhi (discrimination), chitta (memory), and ahamkara (integration) — validated across 2,500+ LLM and multimodal samples. Accompanied by Indian Patent Application No. 202641043947.
Patent protects the system architecture of the Antahkarana cognitively-inspired reasoning engine — specifically the multi-stage query routing mechanism, the Vedantic cognitive stage mapping to neural processing pipelines, and the multimodal integration layer for LLM and VLM joint reasoning. First Indian patent in the space of Vedanta-inspired AI cognitive architectures.
Nine years of high-voltage systems taught me things no AI course ever could. Here's what actually crosses over — and what doesn't.
Most LLM pipelines are glorified prompt chains. Antahkarana is different — it routes queries through Vedantic cognitive stages. Here's the architecture and why it works.
When building Antahkarana's medical reasoning engine, I had to choose between RAG and fine-tuning for every sub-task. Here's the decision framework I developed.
A YouTube channel dedicated to understanding Artificial Intelligence — from fundamentals to advanced concepts like Machine Learning, Deep Learning, NLP, Reinforcement Learning, LLMs, Computer Vision, and Intelligent Agents.
This channel explores AI not just as technology, but as a step toward human-like reasoning, cognition, and consciousness. The same philosophy behind the Antahkarana framework — taught publicly.
"Learn AI deeply. Think consciously."
▶ Visit the Channel →Anbu Health AI is live at anbuclinic.me — serving real patients in Ariyalur District, Tamil Nadu. Invoice AIV-2026-001 for ₹21,500 was issued to Anbu Clinic in June 2026 under AI Vision (Udyam Reg: UDYAM-TN-02-0483528). This is not a demo. It is a paid, production-deployed product with a real client and a maintenance contract.
Most M.Tech graduates build class projects. Rajaganapathy filed an Indian patent (App. 202641043947), submitted an IEEE paper, ran a 2-year freelance AI practice under AI Vision, and deployed a live billed product — all while graduating with a 9.6 CGPA in May 2026.
Nine years maintaining 500 kVA transformers and 400 kW motors in live industrial environments. Then two years building and shipping real AI products for paying clients. The discipline of both is in every line of code — and recruiters can verify it with a live URL and an invoice number.
Most candidates bring a degree. I bring 9+ years of engineering under real-world pressure,
2+ years of freelance AI delivery under AI Vision (Udyam Reg: UDYAM-TN-02-0483528),
a live billed product at anbuclinic.me,
a filed patent, an IEEE paper under review, an ORCID researcher identity,
and a Google Scholar profile — all while completing an M.Tech in AI with a 9.6 CGPA.
You can verify the product right now: anbuclinic.me is live.
CLICK TO VIEW · PRINT · OR DOWNLOAD AS PDF
Share rajaganaa.github.io/?resume=1 with recruiters — auto-opens the resume