AI engineer / researcher / builder

I build AI systems that survive contact with reality.

Agents, retrieval, evaluation and infrastructure. I build them, test them, break them and ship them.

M.S. ARTIFICIAL INTELLIGENCE, RIT · SAN FRANCISCO BAY AREA

Currently building SHAPI Self-maintaining APIs for teams and AI agents. View project →
Photo of Sri Varshini

01 / About me

About me.

Current

Building SHAPI, self-maintaining APIs for teams and AI agents. Recent M.S. graduate in Artificial Intelligence from Rochester Institute of Technology, based in the San Francisco Bay Area.

Background

Bachelor’s in Computer Science from R.M.K. Engineering College, then ML engineering at Novacis Digital and research at RIT, building document intelligence, semantic search and LLM pipelines.

Expertise

AI agents, LLMs, RAG, evaluation, computer vision and cloud-native ML deployment, turning research into systems with measurable impact.

18% F1 improvement Legal document processing
60% Faster lookups Analyst search time
35% Lower cost LLM inference

02 / Education

Education

Rochester Institute of Technology

M.S. in Artificial Intelligence

AUG 2024 — MAY 2026 · ROCHESTER, NY

  • Machine Learning Applications
  • Distributed Systems
  • Engineering Scalable Systems

R.M.K. Engineering College

Bachelor’s in Computer Science and Engineering

AUG 2019 — MAY 2023 · INDIA

  • Computer Science Fundamentals
  • Software Engineering
  • AI/ML Research

03 / Featured projects

Featured projects.

All projects →
Agent infra Currently building

SHAPI

Self-maintaining APIs for teams and AI agents.

  • Maps every third-party call site into an api.lock
  • Fixes kept only when a contract test proves them
  • Opens a pull request with the fix, test and evidence
  • TypeScript
  • Node
  • MCP
  • Claude
Agent product

Nori

Job search infrastructure built around a career graph.

  • Career graph: every resume bullet traces to a fact
  • Fills ATS forms with review-before-submit
  • Drive it from Slack, Telegram, WhatsApp, email or MCP
  • TypeScript
  • Agents
  • MCP
  • Composio
  • Parallel
Multilingual AI

MotherTongue

Translate text, photos and documents into the language you understand best.

  • Text, photo or document in; your language and audio out
  • 150+ language interface, installable PWA
  • Keys stay server-side; routes are rate-limited
  • Next.js
  • TypeScript
  • OCR
  • Speech
Agent product

Chief

A personal agent that notices, proposes, and acts only after you approve.

  • Notices problems across calendar, email and sleep
  • Every write waits for an Approve tap
  • New skills are single files of about 20 lines
  • Python
  • Agents
  • Composio
Agent infra

Kerna

A local-first trust layer for AI agents.

  • Fail-closed permissions for every MCP tool
  • Hard budgets on calls, runtime, cost and output
  • A local receipt for every agent run
  • Rust
  • MCP
  • SQLite
Evaluation

RAGTruth

Can multiple AI agents detect when RAG systems hallucinate?

  • Overall F1 of 0.642 with no training data and no GPU
  • High recall of 0.764 on claim-level hallucinations
  • Task-aware rules drive the largest gain (+0.13 F1)
  • Python
  • LLMs
  • Multi-agent
Computer vision

Video Search

Semantic search across video scenes using natural language.

  • Automatic scene detection with CLIP embeddings
  • Natural-language search over scenes in Qdrant
  • Full-stack app with a FastAPI backend and React frontend
  • CLIP
  • Qdrant
  • FastAPI
Agent infra

AgentOps

A control tower for tracing, debugging and auditing agent runs.

  • Every agent run recorded as a replayable trace
  • Risk classification and human approval for risky steps
  • Demo agent writes a safe runbook for a failed API incident
  • FastAPI
  • React
  • PostgreSQL
  • SQLAlchemy
ML systems

Feature Platform

Real-time feature computation and multi-model serving.

  • Kafka → Feast → Redis streaming feature pipeline
  • Fraud, recommendation and forecasting models share features
  • A/B testing with Ray Serve, monitoring with MLflow
  • Kafka
  • Feast
  • Redis
  • Ray Serve
  • MLflow

04 / Work experience

Work experience.

All experience →

RIT

Graduate Researcher

JAN 2026 — APR 2026 · ROCHESTER, NY

Key achievements

  • Built a zero-shot multi-agent system for detecting hallucinations in RAG outputs.
  • Designed the pipeline that decomposes responses into factual claims, verifies them against source passages, and flags unsupported spans using task-specific decision rules.
  • Evaluated it on RAGTruth across QA, summarization and data-to-text.
  • Multi-agent systems
  • RAG evaluation
  • LLMs
  • Python
Details →

Novacis Digital

Machine Learning Engineer

JUN 2023 — FEB 2024 · INDIA

Key achievements

  • Built a SageMaker-based document AI system with LayoutLM + BERT, improving F1 by 18% on 10K+ legal documents.
  • Deployed an OCR + LLM pipeline on Amazon EKS, cutting review time to under 10 minutes per document.
  • Built a FAISS-based semantic search system with 300 ms retrieval, reducing analyst lookup time by 60%.
  • Benchmarked GPT-4 against Claude for contract analysis, reducing inference cost by 35%.
  • Productionized AWS ML pipelines for batch and real-time workflows.
  • SageMaker
  • LayoutLM
  • BERT
  • EKS
  • FAISS
  • GPT-4
  • Claude
  • AWS
Details →

05 / Technical skills

Technical skills.

Pick a skill to see where I used it.

Languages

Python · used in RAGTruthAgentOpsChiefVideo SearchFeature PlatformDocument IntelligenceGraduate Researcher, RIT

LLMs & agents

Machine learning

Search & data

MLOps & cloud

Backend & web

Also familiar with

  • Git
  • GitHub Actions
  • TensorFlow
  • XGBoost
  • LIME
  • PySpark
  • MongoDB
  • DynamoDB
  • AWS Lambda
  • AWS Bedrock
  • CloudWatch
  • Terraform
  • Snowflake
  • Palantir Foundry

06 / Research

Research publications.

All research →
Publication Published

Lightweight Channel Attention for Efficient CNNs

An empirical study of SE, ECA and a proposed Lite Channel Attention module across ResNet-18 and MobileNetV2 on CIFAR-10, measuring accuracy, parameters, FLOPs and latency.

  • Negligible parameter and latency overhead
  • Competitive accuracy (94.68% on ResNet-18, 93.10% on MobileNetV2)
  • Tested on ResNet-18 and MobileNetV2
Publication Published

EDAT-Lite: Predictive Difficulty-Aware Training for Neural Networks

A lightweight difficulty-aware training framework that learns and predicts hard examples with an auxiliary head, making training dynamics easier to interpret while preserving accuracy on MNIST.

  • Lightweight training framework
  • Improved interpretability of training dynamics
  • Preserved accuracy while enhancing explainability

07 / Certifications

Certifications

Edge AI Specialization

NVIDIA

Edge AI deployment, model optimization and real-time inference.

Machine Learning Specialization

Stanford University

Machine learning algorithms, theory and practical applications.

08 / Get in touch

Get in touch.

Always open to discussing new opportunities, collaborations, or just having a chat about AI and ML.