Machine Learning Engineer
I build intelligent systems that turn complex problems into real solutions — from LoRA/QLoRA fine-tuning and multi-agent LLM pipelines to quantum machine learning and healthcare AI.
Have an ambitious idea?
Let’s build something intelligent.
Clinical LLM pipelines, severity triage, early-warning vitals scoring, FHIR interoperability — AI that helps care teams decide faster.
LoRA/QLoRA fine-tuning, hybrid RAG with rank fusion, multi-agent orchestration, hallucination guards — grounded, citation-checked answers.
Quantum kernels, QSVM, hybrid quantum-classical models — searching for advantage in noisy, low-data clinical regimes.
Reinforcement learning, RLHF, prompt engineering, transfer learning — the training loop end of the stack.
Open a case study for the problem → approach → result breakdown.
A full-stack platform unifying patient records, clinical workflows and ML-assisted decision support — polyglot persistence behind a FastAPI gateway, with a hybrid RAG engine serving citation-checked, hallucination-guarded clinical answers.
Hospital data lives in silos — records, documents, search and retrieval each in different systems, none of them conversational.
Polyglot persistence: PostgreSQL (EHR + RBAC), Elasticsearch (BM25), Milvus (dense vectors), MinIO (documents). Hybrid RAG with reciprocal-rank fusion over a locally hosted Llama 3.1.
Citation-checked, hallucination-guarded clinical Q&A with FHIR R4 interoperability, early-warning vitals scoring, and Docker/Nginx deployment via Cloudflare Tunnel.
A multi-agent LLM pipeline classifying patient symptom descriptions into Mild / Moderate / Severe / Critical tiers with BioMistral-7B & MedAlpaca-13B — reaching 92% accuracy on held-out data.
Free-text symptom descriptions are ambiguous; mis-triage of severe cases is the costliest failure mode.
4-bit QLoRA fine-tune of BioMistral-7B (r=16, α=32); a confidence-aware hard-flag override that escalates safety-critical cases; a ClinicalReasoningAgent for structured explanations.
92% classification accuracy on held-out data, with explainable severity calls and guaranteed escalation on hard flags.
A hackathon AI-governance platform with Need-Role-Level access control — role-based query classification returning GRANTED / DENIED / BLOCKED, backed by a RAG pipeline over 150 synthetic policy documents.
Corporate LLM assistants leak information across roles — access control needs to live inside the AI layer, not around it.
Owned LLM fine-tuning + the FastAPI layer. Built the ChromaDB RAG pipeline; diagnosed overfitting in the fine-tuned model and pivoted to an Ollama system-prompt deployment.
Working NRL access control demo at hackathon pace — an honest engineering pivot that shipped over a fragile fine-tune.
A QML model for ECG disease classification using quantum kernels + QSVM, with a novel physiology-aware quantum feature map and multi-lead entanglement — improving robustness on noisy, low-data ECG.
Classical ECG classifiers degrade fast on noisy signals and small labeled datasets — exactly where clinics live.
Quantum kernels + QSVM with a physiology-aware feature map: multi-lead entanglement encodes inter-lead relationships directly into the circuit.
Improved accuracy and robustness over classical baselines in noisy, low-data regimes.
A neural weight-denoising model that reconstructs corrupted network parameters for an LSTM encoder–decoder translator.
An ElasticNet model predicting biological age from gene-expression data (GSE113957) — scaling, regularization and honest feature engineering on high-dimensional biology.
SDM College. The first proof that hunting for patterns pays off.
Foundation years in Bengaluru — mathematics, ML fundamentals, and the first end-to-end pipelines.
Assistive tech for mobility and independence — the first taste of building for healthcare, and the reason it stuck.
Hackathon pace: scope hard, ship a working demo, defend the engineering decisions in front of judges.
HIS, the multi-agent severity classifier at 92% accuracy, quantum-kernel ECG — research-grade ideas shipped as working systems.
B.Tech expected 2027. Actively seeking AI/ML internships in healthcare and applied machine learning.
BE PART OF IT →Hover any node — the tracer shows which projects it powers.
Judged rooms, timed builds, ranked results.
Placed 4th out of 60 competing teams — build, demo and defend under the clock.
Quadriplegia Assistance System — assistive technology for enhanced mobility and independence.
2nd place at district level — SDM College, grade 12.
Your mail client should now be open with the message pre-filled. If it isn’t, write directly to adyanth.s@gmail.com.