Open to AI/ML Engineer & Data Engineer roles

Vishnu Pulipaka

Software & ML Engineer specializing in end-to-end intelligent systems — from distributed data pipelines and backend infrastructure to LLM evaluation, GPU inference optimization, and low-resource language model research.

2+ yrs experience 13 shipped projects 6 papers & research 200K+ examples evaluated
01

About

Software & ML Engineer with 2+ years of experience across backend infrastructure, LLM evaluation, and inference optimization.

Background

M.S. in Data Science from Montclair State University (GPA 3.87/4.0). Researching low-resource language modeling and LLM fairness alongside engineering work.

Experience
  • Built production backend services and distributed data pipelines, cutting latency and processing time by double-digit percentages
  • Compiled and optimized deep learning models for GPU inference, multiplying throughput with zero accuracy loss
  • Built evaluation frameworks and quality gates that catch model, retrieval, and data regressions before they ship
  • Ran fairness and drift evaluations across LLM releases from multiple providers
What I Can Work On
  • Backend services and distributed data pipelines
  • Model training, evaluation, and inference optimization
  • LLM applications — retrieval, evaluation, and agentic systems
  • Research-grade evaluation of model behavior — fairness, drift, benchmarking
Looking For

Full-time AI Engineer, ML Engineer, or Data Engineer roles where I can own systems end-to-end — from the pipeline to the model to the API serving it.

Location
Dayton, Ohio, USA · Open to relocate
Target roles
AI Engineer · ML Engineer · Data Engineer
Email
Elsewhere
02

Experience

AI/ML Intern — VTS360
Aug 2026 — Present · United States
Software Engineer — 4SS Software Solutions
Jun 2022 — Jul 2024 · India
03

Technical Skills

AI & LLMs
LangChain, RAG, LoRA fine-tuning, prompt engineering, Hugging Face, ChromaDB, Ollama, Sentence Transformers, Whisper, LLM-as-judge, OpenAI, Anthropic/Claude
ML & Training
PyTorch, TensorFlow, scikit-learn, XGBoost, Keras, FinBERT, TensorRT, ONNX, CUDA, SLURM, FSDP, DeepSpeed
Data Engineering
PySpark, Apache Airflow, Snowflake, PostgreSQL, MySQL, Redis, Apache Kafka, Azure Data Factory, Pandas, dbt
Full-Stack
Python, SQL, Java, C++, FastAPI, Flask, Spring Boot, Node.js, React, Next.js, Tailwind, Streamlit
Cloud & MLOps
AWS SageMaker, Lambda, S3, EC2, RDS, Azure, Docker, Kubernetes, MLflow, Optuna, Evidently, Feast, GitHub Actions, Prometheus
04

Projects

AI Engineering
Production RAG Eval PlatformView repo →

RAG evaluation platform benchmarking 3 retrievers against 3 LLM providers — faithfulness, accuracy, latency, cost, and pass rate, with GitHub Actions gates that catch regressions before release.

92% faithfulness · 100% retrieval hit rate · Python, FastAPI, Streamlit, ChromaDB, OpenAI, Anthropic, Ollama
FinSight-AIView repo →

4-agent async financial research system covering US, NSE, and BSE markets with plurality-vote signals, LLM-as-judge scoring, FinBERT sentiment, and Markowitz portfolio optimization.

4 async agents · 5 quality dimensions · Python, FastAPI, FinBERT, ChromaDB, asyncio
VocalyticsView repo →

Voice-to-SQL app converting natural-language queries into guarded SQL and Plotly charts over 3.4M orders and 30M+ line items, with dynamic allowlists and a 29-test eval suite.

3.4M orders · 29 passing tests · Python, Claude, Whisper, PostgreSQL, Plotly
tok-adaptView repo →

Python/CLI library for adapting, extending, pruning, and initializing Hugging Face tokenizers and embedding layers, plus an end-to-end CPT → SFT → DPO pipeline for cross-lingual LLM fine-tuning.

Quantized export · Python, Hugging Face Transformers, CLI
Data Engineering
CDC PipelineView repo →

Exactly-once change data capture pipeline: PostgreSQL WAL → Debezium → Kafka → Flink → Apache Iceberg, chaos-tested against TaskManager kills and network partitions.

50,000 writes/min · Kafka, Debezium, PostgreSQL, Flink, Iceberg
Data LakehouseView repo →

Delta Lake medallion architecture (bronze/silver/gold) with a compaction engine cutting per-partition file count 95% (87 → 4 files) in 14.5s, zero reader downtime.

95% file reduction · PySpark, Delta Lake, Airflow, AWS S3
Data Quality AIView repo →

Real-time data quality and LLM incident-triage platform detecting 6 failure modes, with automated anomaly alerting integrated into CI/CD data pipelines.

HIGH-severity alerts <42ms · Python, LangChain, PostgreSQL, dbt
ML Engineering
TensorRT Inference Optimization EngineView repo →

Compiled ResNet-50 to an FP16 TensorRT 11 engine on CUDA 12.4 with zero accuracy loss across 10K CIFAR-10 images, served via a C++ HTTP server.

3.55× throughput (97→347 FPS) · 4.3× P99 latency cut · TensorRT, CUDA, PyTorch, ONNX
Telco Customer Churn — MLOps PipelineView repo →

XGBoost classifier on 7,043 records with 50-trial Optuna tuning, gated promotion via a two-tier quality gate against the Champion model in the MLflow Registry.

PR-AUC 0.65 · ≤0.01 regression gate · Python, XGBoost, Optuna, MLflow, Evidently, FastAPI
ML Feature Store & Train-Serve Skew DetectionView repo →

Feature pipeline on NYC Taxi data enforcing train-serve parity via PySpark → Feast offline store → Redis. Skew caught at inference time using SHA-256 hash comparison.

Python, PySpark, Feast, Redis, XGBoost, FastAPI
Full-Stack
FlowForge — Workflow Automation EngineView repo →

Visual workflow automation platform: React Flow editor, Spring Boot DAG executor, JWT auth, OAuth credentials encrypted at rest (Jasypt AES-256), live SSE execution logs.

Java, Spring Boot, React, TypeScript, PostgreSQL, Docker
Flash Sale Reservation PlatformView repo →

High-throughput inventory reservations for flash-sale spikes: Redisson Redlock holds, Kafka-backed checkout, live seat availability over WebSocket with polling fallback.

Java, Spring Boot, Redis, Kafka, PostgreSQL, React
Quire — Collaborative WorkspaceView repo →

Real-time collaborative document editor with live multi-cursor editing via Yjs CRDTs, a Tiptap-based React frontend, and a Spring Boot WebSocket hub.

Java, Spring Boot, React, Yjs, Tiptap, PostgreSQL, Redis
05

Research

How Fast Can Reward Models Score? A Systems Study of C++ and PyTorch Inference Runtimes for RLHFarXiv preprint
arXiv:2607.19712 · First author, with Anish Katta and Deva Rohit Reddy Peddireddy

Built a native C++ inference engine on ONNX Runtime for RLHF reward-model scoring and benchmarked it against PyTorch eager mode, torch.compile, and FastAPI serving on CPU and GPU.

  • On CPU, the C++ engine beat every PyTorch baseline by 1.7–1.9×, with non-overlapping confidence intervals; on GPU, torch.compile won at the median and tail.
  • The CPU speedup traced to ONNX Runtime's graph execution, not C++ itself — calling the same ONNX session from Python performed statistically indistinguishably.
Evaluating Dedicated Monolingual and Joint Multilingual Causal Models for Dravidian LanguagesarXiv preprint
arXiv:2608.07727 · Sole author

Trained 5 GPT-2-architecture causal language models from scratch for Telugu, Tamil, Kannada, and Malayalam (4 monolingual + 1 multilingual), benchmarked against XLM-R, mBERT, and mGPT.

  • Dedicated 32K per-language tokenizers beat XLM-R, mBERT, and mGPT on fertility and compression across all four languages, despite a vocabulary 3–8× smaller.
  • Monolingual pretraining beat multilingual pretraining on every language by bits-per-byte, the tokenizer-agnostic metric.
[Title withheld — under double-blind review]Under double-blind review
Release-level fairness drift in large language models · First author

A fairness-regression audit pipeline for detecting whether a new model version introduces fairness regressions, applied to 12 model versions from 5 providers across 7 public bias benchmarks (~200K examples).

  • Stereotype scores held steady or worsened in 67% of version transitions; Gemini-2.5-Pro showed the sharpest increase (Δ = +0.027, p < 0.001).
  • Fairness proved multidimensional: inter-metric correlations were weak (|r| < 0.2) across sentiment, toxicity, and stereotype dimensions.
[Title withheld — under review]Under review
Machine-translated low-resource pretraining corpora · First author

Investigates whether machine translation (IndicTrans2) can generate BabyLM-scale training corpora for low-resource languages, translating the 100M-word English BabyLM corpus into Hindi and Telugu.

  • Translated Hindi/Telugu models reached strong in-domain perplexity but generalized less well than models trained on native, non-translated data.
  • Bilingual English/Hindi and English/Telugu models stayed competitive with larger pretrained multilingual models (Llama-3.2-1B, Sarvam-2B) on syntactic generalization.
MultiModalFedCD: A Privacy-Preserving Federated Learning Framework Combining Capsule Endoscopy and Gene Expression Data for Crohn's Disease DiagnosisIn preparation
Multimodal · Federated learning · Medical AI

A federated learning framework combining capsule endoscopy imaging with gene expression data for privacy-preserving Crohn's disease diagnosis.

Tokenizer Efficiency Across Languages: Telugu, Hindi & EnglishOngoing
First author

Evaluates subword tokenization across Telugu, Hindi, and English to measure token inflation, computational cost, and accuracy-efficiency trade-offs for low-resource, morphologically rich languages.

06

Education & Certifications

Master of Science in Data Science
Montclair State University
GPA 3.87/4.0 · Algorithms & Analysis, NLP, Machine Learning, Big Data Analytics
Sep 2024 — Dec 2025
Montclair, New Jersey
B.Tech, Electronics & Communication Engineering
Sreenidhi Institute of Science & Technology
Nov 2020 — May 2024
Hyderabad, India
AWS Certified Solutions Architect — Associate
Verified
February 2026
07

Let's Work Together

Looking for full-time AI Engineer, ML Engineer, or Data Engineer work. If you're hiring, or just want to talk about any of this, reach out.