OPEN TO AI · ML · DATA ENGINEERING ROLES

Vishnu
Pulipaka

I write code across the ML stack: training and evaluating models, building the pipelines that feed them, and shipping the backend services that serve them. Based in New Jersey, open to relocating.

0
YEARS EXP
0
PROJECTS
0
PAPERS & RESEARCH
0
EXAMPLES EVAL'D
View Projects Get in Touch
SCROLL

Who I Am

I move between AI engineering, ML engineering, data engineering, and backend software, and I don't treat those as separate tracks. The person who writes the ETL pipeline should understand why the model trained on it drifted, and the person who trains the model should know what happens when its API gets hit 1,000 times a second.

On the AI engineering side, I've built a RAG evaluation platform that benchmarks three retrievers against three LLM providers, a four agent financial research system that combines FinBERT sentiment with LLM as judge scoring, and a voice to SQL app that turns natural language into guarded queries over 3.4 million orders. What I actually care about is whether the outputs are correct, not whether the demo looks good.

For ML engineering, I compiled a ResNet 50 model to a TensorRT engine and got a 3.55x throughput gain with zero accuracy loss, and I built an XGBoost churn model with a quality gate that blocks any retrain if it regresses PR AUC by more than 0.01. At Montclair State University I ran a fairness drift study across 12 LLM versions from 5 providers, currently under double blind review. Most recently I coauthored a systems paper benchmarking a C++ ONNX Runtime engine against PyTorch for RLHF reward scoring. The honest finding there was that the speedup came from graph execution, not from C++ itself, which wasn't the answer I expected going in.

The software and data engineering work is where a lot of this actually gets used. I built Kafka based async services and Redis caching at 4SS Software Solutions that cut processing time and API latency by double digit percentages, and I still maintain a Python and SQL pipeline for Saayam For All that has generated over 30,800 validated records so far. I tend to spend more time than most people on the boring parts: schema validation, exception handling on Lambda functions, making sure a pipeline fails loudly instead of quietly.

Open to opportunities
Target Roles
AI Engineer
ML Engineer
Data Engineer
Software Engineer

Technical Expertise

AI & LLMs
LangChainRAGLoRA Fine-tuning Prompt EngineeringHugging Face ChromaDBOllamaSentence Transformers WhisperLLM-as-Judge OpenAIAnthropic / Claude
ML Frameworks & Training
PyTorchTensorFlowscikit-learn XGBoostKerasFinBERT TensorRTONNXCUDA SLURMFSDPDeepSpeed
Cloud & MLOps
AWS SageMakerLambdaS3 EC2RDSAzure DockerKubernetesMLflow OptunaEvidentlyFeast GitHub Actions
Data Engineering
PySparkApache AirflowSnowflake PostgreSQLMySQLRedis Apache KafkaAzure Data Factory Pandasdbt
Backend & Frontend
PythonSQLJava C++FastAPIFlask Spring BootNode.js ReactNext.jsTailwind Streamlit
AI Coding Tools
Claude CodeCursorCodex GitHub Copilot

Where I've Worked

Saayam For All
Data Engineer (Volunteer)
Mar 2026 — Present
United States
4SS Software Solutions
Software Engineer
Jun 2022 — Jul 2024
India

Things I've Built

🔍
AI Engineer
RAG Eval Platform

A RAG evaluation platform benchmarking 3 retrievers against 3 LLM providers. Tracks faithfulness, accuracy, latency, cost, and pass rate, with GitHub Actions gates that catch prompt, retrieval, and metric regressions before release, plus a Streamlit dashboard for monitoring.

92% faithfulness 100% retrieval rate
PythonFastAPIStreamlitChromaDBOpenAIAnthropicOllama
📈
AI Engineer
FinSight-AI

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
PythonFastAPIFinBERTChromaDBasyncio
🎙️
AI Engineer
VocalLytics

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

3.4M orders 29 passing tests
PythonClaudeWhisperPostgreSQLPlotly
🏗️
Data Engineer
Data Lakehouse

End-to-end Delta Lake medallion architecture (bronze/silver/gold) with partition pruning and Z-ordering for scalable analytics pipelines.

PySparkDelta LakeAirflowAWS S3
Data Engineer
CDC Pipeline

Change Data Capture pipeline for real-time database event streaming, with schema evolution handling and backpressure management for reliable downstream delivery.

KafkaDebeziumPostgreSQLPython
🔎
Data Engineer
Data Quality AI

AI-powered data quality monitoring with LLM-assisted rule generation from schema profiles and automated anomaly alerting integrated into CI/CD data pipelines.

PythonLangChainPostgreSQLdbt
🔄
ML Engineer
Telco Customer Churn — MLOps Pipeline

Production-grade churn prediction system: trained an XGBoost classifier on 7,043 records with 50-trial Optuna tuning, gated promotion via a two-tier quality gate enforcing ≤0.01 PR-AUC regression against the Champion model in the MLflow Registry, and automated the lifecycle via GitHub Actions cron, an Evidently drift monitor, a FastAPI endpoint, and a Streamlit dashboard.

PR-AUC 0.65 ≤0.01 regression gate
PythonXGBoostOptunaMLflowEvidentlyFastAPIDocker
🖥️
ML Engineer
TensorRT Inference Optimization Engine

Compiled ResNet-50 to an FP16 TensorRT 11 engine on CUDA 12.4 with zero accuracy loss across 10K CIFAR-10 images; documented an INT8 tactic failure on sm_86 and validated memory via cudaMemGetInfo delta isolation.

3.55× throughput (97→347 FPS) 4.3× P99 latency cut (21.4ms→4.9ms)
TensorRTCUDAPyTorchONNXPython
📊
ML Engineer
ML Feature Store & Train-Serve Skew Detection

An ML feature pipeline on NYC Taxi data that enforces train-serve parity through PySpark. Features are registered in a Feast offline store for training and materialized to Redis for low-latency inference. Serving skew gets caught at inference time using SHA-256 feature-matrix hashes and strict column-order checks.

PythonPySparkFeastRedisXGBoostFastAPI

Publications & Research

How Fast Can Reward Models Score? A Systems Study of C++ and PyTorch Inference Runtimes for RLHF
arXiv 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 a FastAPI serving layer on both CPU and GPU, using repeated independent process launches rather than single run timings.

C++ / ONNX Runtime RLHF Reward Scoring DeBERTa-v3 & Electra 5 Independent Launches per Config
Key Findings
  • On CPU, the C++ engine beat every PyTorch baseline by 1.7 to 1.9 times, with non-overlapping confidence intervals. On GPU it beat eager mode and FastAPI, but torch.compile won at both the median and the tail (p95).
  • The CPU speedup came from ONNX Runtime's graph execution rather than from C++ itself. Calling the same ONNX session from Python performed statistically indistinguishably from the native C++ engine.
  • Batching strategy mattered more than language or runtime choice. Naive padding cost 5 to 8 times the throughput on CPU and 3.5 to 4 times on GPU compared to length aware bucketing.
  • Running separate engine instances per concurrent request scaled worse than a single shared instance, and caused GPU out of memory failures beyond about 4 to 5 concurrent sessions.
[Title withheld — under double-blind review]
EMNLP 2026 — Under Double-Blind Review
Under double-blind review  ·  First author

A release level fairness drift regression audit framework: a reproducible pipeline for detecting whether a new model version introduces fairness regressions before deployment. Applied to 12 model versions from 5 providers across 6 model families and 7 public bias benchmarks (about 200K examples), evaluating sentiment, toxicity, and stereotype proxy dimensions.

12 Model Versions 5 Providers, 6 Families 7 Benchmarks ~200K Examples 1,000 Bootstrap Samples
Key Findings
  • Toxicity showed a small average decline across newer releases, with a mean drift of −0.005, though the 95% confidence interval spanned zero and individual transitions were non-monotonic.
  • Stereotype scores held steady or worsened in 67% of version transitions. Gemini-2.5-Pro showed the sharpest increase, a statistically significant effect (Δ = +0.027, Cohen's d = 0.64, p < 0.001).
  • A robustness ablation that excluded StereoSet and CrowS-Pairs confirmed all findings replicated, with 8 of 8 drift pairs consistent.
  • Fairness proved multidimensional: inter-metric correlations were weak (|r| < 0.2). OpenAI and Meta (LLaMA) showed the lowest stereotype scores, while Gemini showed the highest and most volatile.
[Title withheld — under review]
BabyLM 2026 Workshop — Under Review
Under review at the BabyLM Workshop  ·  First author

Investigates whether machine translation (IndicTrans2, en-indic-1B) can generate BabyLM scale training corpora for low resource languages. Translates the 100M word English BabyLM corpus into Hindi and Telugu, then trains and evaluates monolingual and bilingual GPT-Wee (GPT-2 style) and GPT-BERT models against pretrained multilingual baselines.

English · Hindi · Telugu IndicTrans2 GPT-Wee & GPT-BERT M-BLiMP · SIB-200 · MuBench
Key Results
  • GPT-BERT consistently outperformed GPT-Wee across monolingual and bilingual settings on both perplexity and linguistic benchmarks.
  • Translated Hindi and Telugu models reached strong in domain perplexity, but generalized less well than models trained on native, non-translated CC-100 data when tested on out of domain OpenSubtitles text.
  • Bilingual English/Hindi and English/Telugu models performed comparably despite Hindi's larger resource base, and stayed competitive with much larger pretrained multilingual models (Llama-3.2-1B, Sarvam-2B) on syntactic (M-BLiMP) generalization.
MultiModalFedCD: A Privacy-Preserving Federated Learning Framework Combining Capsule Endoscopy and Gene Expression Data for Crohn's Disease Diagnosis
In 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.

Federated Learning Capsule Endoscopy Gene Expression Medical AI
Monolingual vs. Multilingual Models for Dravidian Languages: A Controlled Comparison Across Telugu, Tamil, Kannada & Malayalam
Ongoing — Target EACL
Target: EACL 2027  ·  First author

A controlled comparison of monolingual vs. multilingual GPT-2 and RoBERTa models across all four Dravidian languages (about 36.4GB of corpus from CC-100, Wikipedia, and Samanantar), evaluated against MuRIL, IndicBERT v2, mBERT, and XLM-R baselines.

4 Languages ~36.4 GB 10 Models 4× A100 GPUs
Tokenizer Efficiency Across Languages: Telugu, Hindi & English
Ongoing — Target MRL
Target: MRL 2026  ·  First author

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

3 Languages Subword Tokenization Token Inflation

Academic Background

🎓
Master of Science in Data Science
Montclair State University
Relevant Coursework: Algorithms & Analysis, Special Topics in NLP, Machine Learning, Big Data Analytics
GPA: 3.87 / 4.0
Sep 2024 — Dec 2025
Montclair, New Jersey
🎓
Bachelor of Technology in Electronics and Communication Engineering
Sreenidhi Institute of Science & Technology
Nov 2020 — May 2024
Hyderabad, India
AWS Certified Solutions Architect — Associate
February 2026
VERIFIED

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.