SOFTWAREDEVELOPER& AI/MLENGINEER

Shipped a production-grade RAG chatbot serving real users on my live portfolio site, with a full ingestion pipeline and a streaming, guardrailed API

• Ingestion: Parsed and chunked the CV PDF into semantic segments (PDF → recursive chunking → 768-dim embeddings → Pinecone upsert)
• Embedding: Embedded text chunks using a transformer model to generate dense vector representations
• Storage & Retrieval: Stored vectors in a Pinecone vector database with top-K retrieval, score thresholding, and metadata filtering to minimize hallucination
• Generation: Built a streaming API endpoint using Vercel AI SDK with Gemini 2.5 Flash for token-by-token delivery, achieving sub-second TTFT
• Security: Hardened the system against prompt injection through input validation, system-prompt design, and topic-scoped guardrails
• Testing: Authored a 27-test integration suite covering embedding quality, retrieval accuracy, pipeline correctness, and injection defense

Built a ReAct tool-use agent that triages production incidents across 5 data sources (RAG over 768 runbooks, text-to-SQL, GCP status API vulnerability DB, web search) using dual Qwen 2.5 models (32B + 3B) on a single GPU via 4-bit quantization.

• Evaluated 4 prompting techniques across 80 benchmarks, finding self-reflection improved answer grounding by 125% (0.40→0.90) and actionability by 100% (0.50→1.0) over baseline, while prompt chaining achieved perfect tool selection on the 32B model but failed on 3B — revealing minimum capability thresholds for structured data
• Achieved ~4% prefill speedup on the 32B model by building KV cache reuse for system prompt acceleration.
• Implemented defense-in-depth security (SQL injection blocking, prompt hardening) passing 3/5 red-team attacks per model; deployed via Gradio with real-time model/technique comparison

End-to-end encoder-decoder image captioning models with attention mechanisms and interactive visualizations.

• Designed and trained three end-to-end encoder-decoder image captioning models (Show-and-Tell baseline, Show-Attend-and-Tell with soft attention, and Visual Sentinel adaptive attention) on MS COCO 2014, leveraging a fine-tuned ResNet-101 CNN encoder feeding a 7×7×2048 spatial feature grid into an LSTM language decoder
• Implemented Bahdanau soft attention with doubly stochastic regularization, producing dynamic per-word context vectors that boosted CIDEr from 1.023 → 1.040 and won all 8 evaluation metrics over the baseline; BLEU-4 of 0.329 materially exceeds the original 2015 paper (0.250)
• Engineered beam search decoding with a 1–10 beam-width ablation that lifted CIDEr from 1.013 (greedy) to 1.093 (beam=6), identifying optimal decoding configuration and quantifying the degradation from overly wide beams
• Built interpretable attention heatmap visualizations that trace each generated word back to the image region driving the prediction — turning a black-box captioner into a debuggable, explainable system
• Optimized training throughput by pre-extracting CNN features to HDF5, achieving full convergence in 30 epochs / ~3 hrs per model on an RTX 6000 with Adam + ReduceLROnPlateau scheduling
• Conducted CIDEr-vs-SPICE discrepancy analysis to expose blind spots in n-gram-only evaluation, demonstrating cases where models score well without semantic understanding
• Deployed as an interactive multi-model demo on Hugging Face Spaces via Docker + Gradio, allowing real-time image upload, model switching, and tunable beam width with live attention heatmaps

FullStack AutoML platform and code generator

• Engineered an AutoML system using Scikit-Learn & Flask that automates preprocessing, task detection, and parallel model training/evaluation
• Designed a responsive React frontend with real-time interactive ROC/Scatter plots and confusion matrix for visualizations of top model
• Reduced training latency by 40% implementing Stratified Sampling and dynamic model switching to handle large datasets efficiently
• Developed a context manager to profile real-time CPU/RAM usage & a leaderboard to sort models by accuracy, time, or resource efficiency
• Built a transpiler engine to enable one-click downloads for both serialized models (.pkl) and their reproduction code

Data analysis and prediction using ML

• Analyzed data of 10,000 account holders at a Multinational Bank by doing exploratory data analysis with Pandas
• Constructed a streamlined pipeline for training 5 machine learning models to predict customer churn using Scikit-Learn
• Implemented XGBoost, Random Forest, KNN, SVM, and Naive Bayes models and compared them
• Utilized N-fold cross-validation, F1-score, confusion matrix to evaluate the performance of each model

CNN-based image classification for skin cancer detection

• Developed a CNN using Scikit-Learn to classify skin lesion images into 7 cancer categories, achieving 80% accuracy
• Preprocessed data with resizing, normalization, one-hot encoding, and oversampling to address class imbalance
• Optimized model performance using the Adam optimizer, learning rate annealing, and hyperparameter tuning