Open to Remote Opportunities

ML Engineer
Building AI that ships.

3 years at PwC building production AI/ML systems for Fortune 500 clients. Specializing in RAG pipelines, multi-agent systems, and LLM applications — from architecture to HuggingFace deployment.

View Projects Get In Touch
3+
Years at PwC
4
Live ML Apps
50TB+
Data Processed
3
Fortune 500 Clients
// projects

Live ML Applications

● Live Multi-Agent
Demo GitHub
SafeRoute AI — Multi-Agent Real-Time Safety System

3-agent orchestration system (ML Predictor · Emergency Responder · Route Analyzer) running in parallel via LangGraph. Predicts failures 48 hours in advance with 99.8% accuracy, executes automated 5-step emergency protocol, and maintains sub-500ms response latency end-to-end.

99.8%failure prediction accuracy
48 hrsadvance prediction window
<500msend-to-end latency
LangGraph ThreadPoolExecutor Random Forest Ollama Groq Streamlit
● Live RAG · LLM
Demo GitHub
RAG Agent System — Production Document AI

End-to-end multi-PDF RAG pipeline with FAISS vector search across 10K+ chunks. Features a 3-tier LLM fallback chain (Groq → OpenAI → FLAN-T5), source attribution, chat history, and local hosting with no API key required. 10× faster than baseline.

10K+chunks indexed
~1sresponse time
10×faster than baseline
LangChain LCEL LangGraph FAISS Groq API Sentence-Transformers Gradio FastAPI
● Live Analytics
Demo GitHub
LLM Comparison Dashboard

Benchmarked 3 GPT-2 variants across 5 NLP metrics with 40% faster inference via ThreadPoolExecutor parallelism. DistilBERT achieves 89% accuracy. All 100+ benchmark sessions tracked persistently in SQLite with interactive Plotly visualizations.

40%faster inference
89%DistilBERT accuracy
100+sessions tracked
Streamlit HuggingFace Transformers GPT-2 DistilBERT Plotly SQLite
🔒 Internal Tool Code Analysis · PwC
SAP ABAP Code Analyzer

AI classification pipeline using LangChain + GPT-4o with a deterministic override layer to assess legacy ABAP codebases for SAP S/4HANA migration. Engineered a 4-level token-aware fallback system — reducing consultant assessment time from weeks to minutes.

Weeks → minsassessment time reduction
4-leveltoken-aware fallback
Zeropipeline failures
Python LLM SAP ABAP Streamlit Gradio
// skills

Technical Stack

LLM & RAG
LangChain LangGraph RAG Pipelines FAISS Groq Ollama Prompt Engineering
ML & Deep Learning
PyTorch Scikit-learn Random Forest HuggingFace Transformers CNNs Transfer Learning
Deployment
HuggingFace Spaces Streamlit Gradio FastAPI Docker GitHub Pages
Data Engineering
SAP Migration Syniti BODS Alteryx SQL ETL Pipelines LTMC
// experience

Where I've Worked

Data & ML Engineer
@ PwC India
Mar 2023 — Present
  • Built an AI classification pipeline using LangChain + GPT-4o with deterministic override layer to assess legacy ABAP codebases for SAP S/4HANA migration — reducing consultant assessment from weeks to minutes
  • Engineered a 4-level token-aware fallback system (full context → RAG retrieval → chunked summarization → LLM merge) using tiktoken + LangChain — ensuring zero pipeline failures on codebases of any size
  • Designed a dual RAG architecture using FAISS + OpenAI Embeddings — global store for cross-project retrieval and scoped per-module stores — auto-generating technical and functional spec docs as Word/Excel outputs
  • Diagnosed and resolved silent production upload failures caused by three compounding issues (middleware conflict, CORS preflight mishandling, read-only container FS) — restoring full functionality with zero data loss
  • Reduced LLM API costs via input-hash-based call caching, model tiering, and batched file classification — eliminating redundant API calls on repeat runs
BioMarin Kimberly-Clark Rockwell Automation
// contact

Let's Build Something

📧
email
pravallipasala@gmail.com
💼
linkedin
pasala-pravallika
🐙
github
Prav-allika
🤗
huggingface
Prav04