• About Me
• Featured Projects
• Tech Stack
• GitHub Stats
• Currently
• Get in Touch

class Mahateer:
    def __init__(self):
        self.role = "AI/ML Engineer"
        self.focus = ["LLM Systems", "Preference Optimization", "Applied Deep Learning"]
        self.background = "Co-founder @ UXELERATE, university student"
        self.currently_learning = "LLM/RAG, applied AI systems"
        self.fun_fact = "trained a DPO model on a single T4 in ~2 hours"

    def say_hi(self):
        print("Let's build something that doesn't hallucinate.")

I build production-grade AI systems — LLM orchestration with real security hardening, preference-optimization fine-tunes, clinical data pipelines, and interactive ML tooling. Currently deepening into LLM/RAG and applied AI systems.

A single API call, 25 nested LLM calls, and enough hardening to survive contact with the real world

FastAPI service where /api/v1/run spawns up to 25 nested LLM calls across Proponent/Opponent agents. Pushed through a full LLMOps + security hardening pass:

• 🛡️ Distributed rate limiting (slowapi + Redis, cross-worker/pod safe)
• 🛡️ Prompt injection defense — XML delimiter encapsulation, input tag-stripping, system-level security directive
• 📊 Request tracing with correlation IDs, structured JSON logs
• 🔒 Docker hardened — non-root user, cap_drop: ALL, correct ownership
• ⚡ Async connection pooling moved outside the retry loop, asyncio.wait_for timeouts

Teaching a 7B model to stop confidently making things up in clinical answers

Fine-tunes Qwen2.5-7B-Instruct with Direct Preference Optimization (not SFT) on the MedHallu benchmark — matched grounded vs. hallucinated PubMedQA answers, 4-bit QLoRA, 40.3M trainable LoRA params (0.53% of total weights), ~2 hours on one Kaggle T4.

Benchmark results (999 held-out human-annotated pairs):

Hard-tier score (0.586) lands close to GPT-4o's reported ~0.625 on the same split — from a model 1/20th the size, fine-tuned on a free GPU.

Case study — the model catching a fatal contraindication:

Query: Should beta-blockers be given immediately in cardiogenic shock from acute MI?

❌ Base model: "Yes, initiate immediately in all AMI patients." (dangerous — beta-blockers are contraindicated here) ✅ MedTrust: "No — contraindicated in cardiogenic shock due to negative inotropic effects. Stabilize hemodynamics first."

Neural networks, explained by the network itself

• 🖼️ Network Canvas — semantic-zoom D3 graph, VCR-style stepping
• 🧪 Activation Lab, CNN Lab (filter factory, receptive fields, saliency), Optimizer Arena, BatchNorm Tracker
• ✅ 255 tests, zero faked data — every number comes live from PyTorch hooks

40GB of raw hospital data in, a risk-scoring RAG agent out

Production pipeline processing 546K+ MIMIC-IV hospital stays (40GB+ raw temporal tables) into ML-ready Parquet datasets, feeding multi-task clinical prediction models. This is the project MedTrust's hallucination-reduction work feeds into — same failure mode, applied to the decision-support agent below.

5-Stage Pipeline: Load (schema inference) → Clean (validation) → EDA → Feature Engineering → Datasets

• 🎯 0.949 AUROC — 24-hour mortality prediction
• 🎯 0.897 AUROC — 6-hour ward deterioration
• 🧩 LLM/RAG decision-support agent: SHAP TreeExplainer interpretability, counterfactual "what-if" simulation, clinical guideline retrieval (KDIGO, Surviving Sepsis)
• 🔗 Patient embedding layer for similar-patient retrieval

Distributed retail analytics on PySpark — revenue analysis, customer metrics, spend classification, rolling averages.

Healthcare AI that shows its work

• 🎯 90.69% accuracy on 4-class tumor classification (ResNet50 transfer learning)
• 📊 Grad-CAM explainability layered on top of every prediction
• 🎨 Interactive Streamlit dashboard
• 🧪 Five-notebook pipeline: EDA → Preprocessing → Custom CNN → Transfer Learning → Grad-CAM

Two Flutter apps from earlier in my dev work — real-time systems with Firebase backends:

• Social-Swap (Konexea) — social app with real-time chat, AI-assisted translation, Rive animations
• CALiNGA — on-demand healthcare platform with live location tracking and provider matching

LLM Systems & Orchestration   ████████████████░░ 85%
AI/ML Development             █████████████████░ 90%
Data Engineering               ██████████████░░░░ 80%
Backend Development            █████████████░░░░░ 85%
Frontend Development           ████████████░░░░░░ 75%
DevOps & Deployment              ████████░░░░░░░░░░ 45%

• 🔭 Working on preference-optimization methods (DPO) for factual grounding in LLMs
• 🌱 Learning applied RAG architectures and production LLMOps
• 🤝 Open to collaborating on applied LLM and systems projects
• ⚡ Fun fact: trained MedTrust's DPO adapter (40.3M params) end-to-end on a free Kaggle T4 in under 2 hours