python3 — ml_engineer.py
ANISH
KUMAR
BTech AI/ML — Final Year
Lovely Professional University
I write ML code that survives git push --force.
Final-year BTech AI/ML student building production-ready machine learning systems. I don't just train models — I deploy them. My focus is the full ML lifecycle: data pipelines, training, inference optimisation, API serving and monitoring.
Currently building local LLM inference servers with vLLM, quantised model deployment with GGUF, and RAG evaluation pipelines with RAGAS metrics.
- → Quantised Mistral server
- → RAG eval with RAGAS
- → Edge inference w/ GGUF
Tools I work with
Checkpoints.
“Structured feedback that improved prompt acceptance by 18%.”
Projects
BERT Sentiment Classifier
Fine-tuned a BERT model for five-class sentiment analysis and deployed it as a containerized FastAPI service. Optimized inference with ONNX and tracked experiments using MLflow, enabling scalable real-time predictions through REST endpoints.
Churn Prediction Pipeline
End-to-end churn predictor with SMOTE, SHAP explainability and a real-time Streamlit dashboard.
Document Q&A — RAG
RAG pipeline for multi-document Q&A with FAISS vector search and transformer embeddings.
MedReport AI — Medical Report Analyzer
A 4-stage pipeline that turns raw lab-report PDFs/scans into plain-language explanations using PaddleOCR, Regex/BioBERT NER, anomaly scoring against a 50+ lab-test database, and Gemini API for explanation generation.
Predictive OS Resource Allocation
A smart OS resource allocator that predicts process behavior and optimizes CPU & memory utilization using dynamic scheduling and real-time analytics.
JAS — Natural-Language Agent
An LLaMA 3.1-powered agent that turns natural-language instructions into desktop and browser automation actions.
AI-Powered Medical Report Analyzer
The Challenge
Medical lab reports are dense, jargon-heavy documents — a printed PDF full of values, units, and reference ranges that mean nothing without a clinical background. The goal was to build an end-to-end pipeline that extracts structured data from raw report scans and turns it into a plain-language explanation, without requiring a clinician in the loop for the first pass.
The Architecture
PaddleOCR converts uploaded PDFs/images into raw text, handling both scanned and digitally-generated report layouts.
Regex-based fuzzy matching pulls structured lab values. BioBERT NER pipeline included for messier, non-templated reports.
Values compared against a reference DB of 50+ common lab tests (CBC, LFT, Lipid, etc.) and scored by deviation severity.
Anomalies passed to an LLM (Gemini) with a constrained prompt to generate a patient-friendly summary.
Key Results
- 95%+ Extraction AccuracyAcross diverse test report layouts.
- 50+ Lab Tests SupportedIncluding CBC, LFT, KFT, Lipid, Thyroid, and Glucose.
- Sub-2s LatencyEnd-to-end from PDF upload to full LLM response.
