Rituraj Kulshresth

AC

Enterprise platform for designing, deploying, and operating AI agents. Designed the four-resource credentials model (Team / App / Agent / Sub-agent) with separate edit and invocation ACLs, dual auth modes (act_as_app via API keys + AWS Secrets Manager, act_as_user via Okta OIDC), and a sub-agent safety property where only the caller's identity flows through every tool call - making cross-team agent reuse safe by construction. Refactored agent configuration off YAML onto a versioned PostgreSQL JSONB registry; shipped multi-model serving via AWS Bedrock alongside OpenAI and Anthropic. Drove SAST findings from CRITICAL to zero in a single sprint.

Kn

Self-serve retrieval service on AWS OpenSearch where teams create their own enterprise document indices and share both the document and index catalogs across teams - so any team building an AI app can reuse existing indices without re-ingesting. Ingestion from Confluence, Jira, GitHub, S3, Airflow logs, and arbitrary text sources, with bulk-upload, team-specific chunking, and a semantic chunking mode. Multimodal embedding support via AWS Bedrock Titan, Anthropic Claude, and a Bedrock multimodal model for diagrams, screenshots, and scanned PDFs. Authored the org-wide architecture document.

Da

Internal multi-agent platform used by 11+ engineering and business teams to build domain-specific chatbots with fine-grained control over tools, data sources (SQL, text), and model configurations. Cut monthly infra spend from ~$700 to ~$200 per non-prod environment and ~$1000 to ~$800 in PROD. Built a DeepEval-based evaluation framework that improved evaluation accuracy by 27%, and reduced chat-session storage by ~66% by optimising how Vercel AI SDK logs sessions into OpenSearch. Migrated all Lambdas onto the org's standardised IaC accounts with zero downtime.

S2

Microsoft Outlook add-ins that ship documents from email into enterprise archival systems. S2SACC moves docs to ContentCloud (1000+ users) using Yeoman and MS Graph API with role-based access via distribution groups; S2EDS sends to enterprise servers (500+ users) using React. Drove the Mule ESB → AWS Lambda migration for the archival backend, reducing compute cost while maintaining four critical Java applications.

Ru

Published a clipboard manager that gives Debian-based distros convenient access to clipboard history. Written in Rust, packaged for everyday workflow use.

x8

Hand-written 32-bit OS in C and Assembly with keyboard input, a VGA output driver, and a functional shell for basic tasks. The project that taught me to read datasheets.

RG

CNN models in TensorFlow that predict core body temperature from diverse facial image data (NIR, hyper-spectral, RGB), achieving 60% accuracy on a custom dataset.

Ed

Reduced data transferred between vehicles and Road-Side Units using edge-based knowledge inference. Built on NS-3 with SUMO traffic simulation.

Ha

Modelled how hashtags spread across Twitter and Koo - graph structure, cascade depth, and influencer reach. Useful primer in network science.

Bl

Turns an Android phone into a wireless mouse via Bluetooth - pointer, scroll, and click. My first end-to-end shipped app.

Built a graph-based structure that uses article viewership counts to surface anomalies across Wikipedia articles, improving detection efficiency over naive baselines.

EC

A MATLAB application for ECG signal analysis: heart-rate detection and classification of arrhythmia types, built during a biomedical instrumentation course.

An application that solves Kakuro puzzles by modelling the grid as a constraint-satisfaction problem and searching for valid digit assignments.

AC

Enterprise platform for designing, deploying, and operating AI agents. Designed the four-resource credentials model (Team / App / Agent / Sub-agent) with separate edit and invocation ACLs, dual auth modes (act_as_app via API keys + AWS Secrets Manager, act_as_user via Okta OIDC), and a sub-agent safety property where only the caller's identity flows through every tool call - making cross-team agent reuse safe by construction. Refactored agent configuration off YAML onto a versioned PostgreSQL JSONB registry; shipped multi-model serving via AWS Bedrock alongside OpenAI and Anthropic. Drove SAST findings from CRITICAL to zero in a single sprint.

Kn

Self-serve retrieval service on AWS OpenSearch where teams create their own enterprise document indices and share both the document and index catalogs across teams - so any team building an AI app can reuse existing indices without re-ingesting. Ingestion from Confluence, Jira, GitHub, S3, Airflow logs, and arbitrary text sources, with bulk-upload, team-specific chunking, and a semantic chunking mode. Multimodal embedding support via AWS Bedrock Titan, Anthropic Claude, and a Bedrock multimodal model for diagrams, screenshots, and scanned PDFs. Authored the org-wide architecture document.

Da

Internal multi-agent platform used by 11+ engineering and business teams to build domain-specific chatbots with fine-grained control over tools, data sources (SQL, text), and model configurations. Cut monthly infra spend from ~$700 to ~$200 per non-prod environment and ~$1000 to ~$800 in PROD. Built a DeepEval-based evaluation framework that improved evaluation accuracy by 27%, and reduced chat-session storage by ~66% by optimising how Vercel AI SDK logs sessions into OpenSearch. Migrated all Lambdas onto the org's standardised IaC accounts with zero downtime.

S2

Microsoft Outlook add-ins that ship documents from email into enterprise archival systems. S2SACC moves docs to ContentCloud (1000+ users) using Yeoman and MS Graph API with role-based access via distribution groups; S2EDS sends to enterprise servers (500+ users) using React. Drove the Mule ESB → AWS Lambda migration for the archival backend, reducing compute cost while maintaining four critical Java applications.

Ru

Published a clipboard manager that gives Debian-based distros convenient access to clipboard history. Written in Rust, packaged for everyday workflow use.

x8

Hand-written 32-bit OS in C and Assembly with keyboard input, a VGA output driver, and a functional shell for basic tasks. The project that taught me to read datasheets.

RG

CNN models in TensorFlow that predict core body temperature from diverse facial image data (NIR, hyper-spectral, RGB), achieving 60% accuracy on a custom dataset.

Ed

Reduced data transferred between vehicles and Road-Side Units using edge-based knowledge inference. Built on NS-3 with SUMO traffic simulation.

Ha

Modelled how hashtags spread across Twitter and Koo - graph structure, cascade depth, and influencer reach. Useful primer in network science.

Bl

Turns an Android phone into a wireless mouse via Bluetooth - pointer, scroll, and click. My first end-to-end shipped app.

Built a graph-based structure that uses article viewership counts to surface anomalies across Wikipedia articles, improving detection efficiency over naive baselines.

EC

A MATLAB application for ECG signal analysis: heart-rate detection and classification of arrhythmia types, built during a biomedical instrumentation course.

An application that solves Kakuro puzzles by modelling the grid as a constraint-satisfaction problem and searching for valid digit assignments.

AC

Enterprise platform for designing, deploying, and operating AI agents. Designed the four-resource credentials model (Team / App / Agent / Sub-agent) with separate edit and invocation ACLs, dual auth modes (act_as_app via API keys + AWS Secrets Manager, act_as_user via Okta OIDC), and a sub-agent safety property where only the caller's identity flows through every tool call - making cross-team agent reuse safe by construction. Refactored agent configuration off YAML onto a versioned PostgreSQL JSONB registry; shipped multi-model serving via AWS Bedrock alongside OpenAI and Anthropic. Drove SAST findings from CRITICAL to zero in a single sprint.

Kn

Self-serve retrieval service on AWS OpenSearch where teams create their own enterprise document indices and share both the document and index catalogs across teams - so any team building an AI app can reuse existing indices without re-ingesting. Ingestion from Confluence, Jira, GitHub, S3, Airflow logs, and arbitrary text sources, with bulk-upload, team-specific chunking, and a semantic chunking mode. Multimodal embedding support via AWS Bedrock Titan, Anthropic Claude, and a Bedrock multimodal model for diagrams, screenshots, and scanned PDFs. Authored the org-wide architecture document.

Da

Internal multi-agent platform used by 11+ engineering and business teams to build domain-specific chatbots with fine-grained control over tools, data sources (SQL, text), and model configurations. Cut monthly infra spend from ~$700 to ~$200 per non-prod environment and ~$1000 to ~$800 in PROD. Built a DeepEval-based evaluation framework that improved evaluation accuracy by 27%, and reduced chat-session storage by ~66% by optimising how Vercel AI SDK logs sessions into OpenSearch. Migrated all Lambdas onto the org's standardised IaC accounts with zero downtime.

S2

Microsoft Outlook add-ins that ship documents from email into enterprise archival systems. S2SACC moves docs to ContentCloud (1000+ users) using Yeoman and MS Graph API with role-based access via distribution groups; S2EDS sends to enterprise servers (500+ users) using React. Drove the Mule ESB → AWS Lambda migration for the archival backend, reducing compute cost while maintaining four critical Java applications.

Ru

Published a clipboard manager that gives Debian-based distros convenient access to clipboard history. Written in Rust, packaged for everyday workflow use.

x8

Hand-written 32-bit OS in C and Assembly with keyboard input, a VGA output driver, and a functional shell for basic tasks. The project that taught me to read datasheets.

RG

CNN models in TensorFlow that predict core body temperature from diverse facial image data (NIR, hyper-spectral, RGB), achieving 60% accuracy on a custom dataset.

Ed

Reduced data transferred between vehicles and Road-Side Units using edge-based knowledge inference. Built on NS-3 with SUMO traffic simulation.

Ha

Modelled how hashtags spread across Twitter and Koo - graph structure, cascade depth, and influencer reach. Useful primer in network science.

Bl

Turns an Android phone into a wireless mouse via Bluetooth - pointer, scroll, and click. My first end-to-end shipped app.

Built a graph-based structure that uses article viewership counts to surface anomalies across Wikipedia articles, improving detection efficiency over naive baselines.

EC

A MATLAB application for ECG signal analysis: heart-rate detection and classification of arrhythmia types, built during a biomedical instrumentation course.

An application that solves Kakuro puzzles by modelling the grid as a constraint-satisfaction problem and searching for valid digit assignments.