Porumamilla Akash

Production RAG application for multi-document PDF question answering, conversational retrieval, grounded responses, and source attribution.

• Built and deployed an end-to-end RAG application for multi-document PDF question answering with grounded responses and source/page attribution.
• Designed a LangGraph workflow that rewrites conversational follow-up questions, retrieves relevant chunks through MMR-based FAISS search, and generates answers using Groq GPT-OSS 120B.
• Implemented PDF ingestion, recursive chunking, FastEmbed embeddings, FAISS indexing, duplicate-upload protection, and conversational history.
• Built FastAPI endpoints for document management, PDF upload/indexing, and conversational inference, with a separately deployed Streamlit frontend.

Retrieval-grounded LLM agent for support-ticket triage with intent classification, grounded reply drafting, deterministic escalation, and evaluation.

• Classified support tweets into 10 data-derived intents, retrieved historical resolutions, drafted grounded replies, and routed auto-handle versus escalation in a structured LLM workflow.
• Designed deterministic escalation guardrails that prevent prompt-injected customer messages from bypassing safety decisions.
• Built a 200-example stratified evaluation set with two baselines, bootstrap confidence intervals, and LLM-as-judge evaluation; achieved 74.5% auto-handling at 0.91 precision.

Clinical software platform for chronic kidney disease risk staging, longitudinal monitoring, alerting, and doctor–patient communication.

• Built a 60-endpoint FastAPI backend, roughly 6,100 lines of Python, over 17 PostgreSQL tables with Alembic migrations.
• Trained a Gradient Boosting classifier staging CKD risk into four KDIGO-aligned levels from five clinical biomarkers plus five engineered features.
• Implemented WebSocket doctor–patient chat, push notifications, medication reminders, and PDF/OCR extraction of lab reports.
• Shipped the application to Android and iOS through Capacitor wrapping the deployed Next.js frontend.

Final-year B.Tech thesis at IIT Bhilai: a replication of AFRAIDOOR (IEEE TSE 2024), a data-poisoning backdoor attack on neural code models.

• Reproduced fixed, grammar-sampled, and adaptive gradient-based trigger attacks on an LSTM sequence-to-sequence model trained on CodeSearchNet-Python; the fixed and grammar triggers reached 98.1% and 96.5% attack success at 10% poisoning while clean-task F1 held steady.
• Fixed 9 blocking defects in the upstream research code, including an encoder architecture mismatch between the training and attack code paths and breaking API changes in PyTorch and tree-sitter.
• Cut adaptive trigger generation from 140 to 5.4 seconds per batch (26x faster) by measuring the true maximum identifier count and correcting an over-provisioned hyperparameter, with no loss of coverage.

End-to-end batch data pipeline from PostgreSQL through PySpark and dbt, orchestrated by Airflow and validated in CI.

• Moved 500,000 synthetic shipment records from PostgreSQL into partitioned Parquet with PySpark, using broadcast joins against three dimension tables and window functions for per-customer sequencing.
• Modelled the warehouse layer in dbt across 6 models with 11 schema tests, orchestrated Spark, dbt run, and dbt test as a 3-task Airflow DAG, and added GitHub Actions CI on every push.

Full-stack restaurant application with a React and TypeScript frontend, Express API, transactional email, validation, and rate limiting.

• Built a nine-page restaurant site in React and TypeScript using Vite, shadcn/ui, Tailwind CSS, and a Three.js 3D hero section.
• Wrote an Express API serving reservation and contact submissions, delivering transactional email through the Brevo API with automatic retry on transient failures.
• Added a per-IP sliding-window rate limiter and server-side validation to protect public form endpoints from abuse.