AI product management · Engineering

Building AI tools that make work easier.

Hi, I'm Shrishti. I enjoy understanding how people work and finding practical ways technology can help. Experience in automation at Accenture and AI development at two startups has shaped my interest in AI product management.

At SCU's AIM Lab, we're exploring how AI can support dementia caregivers through MedicalQA. Alongside research, I help organize the NextGen AI4Medicine Workshop.

MS in Engineering Management and Leadership · Santa Clara University · AI and computer science focus · Graduating December 2026

Santa Clara, CaliforniaSeeking AI PM, APM and RPM roles

View experience GitHub LinkedIn

4+ yearsAutomation, stakeholder work, and team leadership at Accenture

Two startupsFintech at Blaze and consumer AI at The Experts Tribe

AIM LabOngoing MedicalQA research for dementia caregivers

25+ professionalsFeedback on DevTwin through conversations and a survey

Featured work

A closer look at my work.

Selected work in AI systems, research, and product development, with the decisions and questions behind each project.

Blaze · AI knowledge system

Second Brain

Second Brain connects Blaze's code, Slack conversations, documents, and meeting notes in one wiki. My role covered designing the architecture and building the system, including how information was organized and kept current.

To control cost, I proposed choosing models according to the content's needs, setting spending limits, and refreshing only what had changed.

Connected contextCode and business information in one wiki

Cost controlModel selection, spending limits, and updates only when needed

Experience details

AIM Lab · Ongoing research

MedicalQA for Dementia Caregivers

At AIM Lab, we're studying how AI can turn educational videos into clear, useful Q&A for dementia caregivers. Our work brings together answer generation, retrieval, and human evaluation.

The current study includes 225 Q&A pairs across five quality dimensions. With source videos available alongside each answer, we can examine quality in context.

4 approachesComparing individual agents, teams of agents, and retrieval

225 Q&A pairsThe current set being reviewed by human evaluators

Research details

Academic research · NLP and model evaluation

Beyond Flesch

We studied how to estimate the difficulty of educational text, comparing a baseline built from 63 prompts with ELECTRA, DANN, and Phi 3.5 Mini adapted using LoRA.

In our experiments, the compact LoRA model offered the strongest overall balance of predictive quality and computing cost.

0.872Test macro F1 for the LoRA model

127×Faster inference than the baseline built from prompt features

Research GitHub Hugging Face Architecture Demo Paper PDF

Academic project · Hardware and human factors

ML Wrist Exoskeleton

We built a passive wrist support prototype using 3D printing, sensor data, and machine learning. Alongside the build, I proposed statistical analysis to assess whether the design actually helped users.

3D printingA physical support structure and straps

EvaluationSensor data, classification, and a plan to assess user benefit

GitHub

Experience

My experience so far.

Current role

Santa Clara University · AIM Lab

Research Assistant

MedicalQA research · Part time · Santa Clara, CA

Jun 2026 to present

Our MedicalQA research explores how questions and answers generated from educational videos could support dementia caregivers.
We compare Qwen generation using one, two, and multiple agents, alongside RAG with BGE embeddings for retrieval.
My contribution includes generation experiments and support for the evaluation interface. Human evaluators review 225 Q&A pairs across five quality dimensions, with access to the source videos.
We support saved ratings, Supabase storage, retries, and result exports so evaluators can return to their work and the team can review the findings.

MedicalQAQwenRAGBGE embeddingsHuman evaluationSupabase

Previous role

Blaze (YC S24)

Special Projects Intern · Core Engineering Team

Reported directly to the CTO and CEO

Early stage fintech startup · International payments · Remote

Jun 2026 to Aug 2026

Designed and implemented Second Brain's architecture, including its connected wiki, diagrams, knowledge graph, and MCP access.
Proposed model choices and spending controls to reduce unnecessary processing, then automated updates through Cloud Run and GitHub Actions.
Built transaction categorization and statement reconciliation in TypeScript and NestJS, with confidence checks, previews, and regression tests for matching and amount errors.
Integrated Puzzle, QuickBooks, and Xero using OAuth and GraphQL, exposing accounting reports to AI CFO tools with separate checks for financial changes.
Created a workflow that converts invoice emails into draft bills, with duplicate and accuracy checks. The code was merged; production activation was still pending when documented.

AI architectureModel cost decisionsTypeScriptNestJSGraphQLMCPGCP

Previous role

The Experts Tribe

AI/ML Engineer Intern

Preseed startup · YOKE dating app prototype · Remote

May 2026

Contributed to YOKE, a dating app prototype for entrepreneurs, across its FastAPI, React, and TypeScript components.
Compared models for frequent writing requests and migrated the prototype from Claude to Gemini. Venue recommendations moved from Yelp to Google Places.
Developed autocomplete, grammar correction, tone rewriting, conversation starters, and venue suggestions, with caching and controls to avoid repeated calls.
Added location consent and a Wingman assistant limited to public profile information. Tests checked that private chats and questionnaire responses stayed out of the model context.

PythonFastAPIReactGeminiGoogle PlacesPrivacyUnit testing

Previous role

Accenture

Senior Analyst · Analyst · Associate

Enterprise software automation · BT EE · India and UK

Jul 2019 to Feb 2024

Senior Analyst, December 2021 to February 2024: worked with BT EE stakeholders to clarify requirements, choose automation tools, and prioritize delivery for a team of five across India and the UK.
Helped lead TEDRepo, an internal AI Q&A chatbot, contributing use cases, KPIs, and conversation flows. The work also included a CI/CD pipeline for daily AI summaries, with solutions adopted across three BT divisions.
Analyst, May to November 2021: developed Bill Clearance automation in Python and Java, reducing manual effort by 90% and cost by 75%.
Associate, July 2019 to April 2021: helped establish the automation practice and expand SAP, CRM, and payment testing to more than 1,500 scripts, with 90% regression coverage.
Earned two promotions in one year and mentored engineers as the team grew.

Stakeholder requirementsTool selectionPrioritizationTeam leadershipPythonJavaCI/CD

Current role

NextGen AI4Medicine Workshop

Workflow Coordinator · Organizing Team

Volunteer · Organizing team

Aug 2026 to present

As part of the organizing team, I help with the workshop website and participant registration form.
We use shared spreadsheets to track tasks, owners, deadlines, and progress. My role includes setting priorities and following up on work that needs attention.
We're preparing for the November 14 and 15 workshop and MediHack on the second day, where students will build and present ideas for AI in medicine.

Organizing teamWebsite and registrationPrioritizationPlanning spreadsheetsTimeline tracking

Visit workshop website

More projects

Things I've built and explored.

These include work from my startup roles, independent projects, and ideas developed through coursework.

Blaze · Accounting integrations

AI CFO & Reconciliation

The reconciliation workflow combines Gemini extraction with matching rules to flag missing, duplicate, or inconsistent records. Connected reports from three accounting providers give the AI CFO a consistent view of the data.

OAuth 2.0GraphQLGeminiJest

Role details

The Experts Tribe · Consumer AI

YOKE Chat Assistant

The YOKE assistant offers writing suggestions, conversation starters, and venue recommendations. The work also covered location consent and checks to keep private information out of the Wingman assistant.

FastAPIReactGeminiGoogle Places

Role details

Independent project · AI for job seekers

AI Career Launchpad

Career Launchpad brings job matching, portfolios, outreach, skill practice, and career Q&A into one app. The design started with personas, product requirements, and KPIs.

Five workflows connect React, Express, FastAPI, ChromaDB, NVIDIA NIM, and live Adzuna data to support different parts of the job search.

Product conceptPRD and KPIsRAGFastAPIReact

GitHub Demo

Independent project · Audio and AI safety

Agent Shield Audio

This FastAPI service transcribes and translates audio, then checks for instructions that could redirect an AI agent. It combines Faster Whisper, Prompt Guard, and keyword checks to flag potential prompt injection.

FastAPIFaster WhisperPrompt GuardPrompt injection checks

GitHub

Academic project · Technology entrepreneurship

DevTwin

DevTwin began as my idea for an AI assistant that helps developers report their work. Conversations and a Google Forms survey brought feedback from more than 25 professionals, including NVIDIA engineering managers and people at Microsoft and Apple.

We used those insights to refine the concept, develop a Figma prototype, explore pricing and positioning, and present the idea to investors as part of our Technology Entrepreneurship course.

User discoveryFigma prototypePricingInvestor presentation

Prototype

Academic project · Product strategy

GridSense

For this smart commercial lighting concept, we explored who would use it, what they might pay, and how to bring it to market. Our coursework covered customer personas, market sizing, KPIs, pricing, and distribution.

Market sizingPricingSaaSLaunch planning

Discuss details

Academic project · Speech and learning

Accent2Edu

We combined speech recognition, keyword extraction, and video search to recommend learning videos from spoken English, including accented speech. The goal was to make relevant educational content easier to find.

NVIDIA NeMoKeyBERTPythonYouTube API

GitHub

Learning resource · Interactive notes

Deep Learning Study Guide

Created for deep learning revision, this guide combines interactive notes, formulas, and lecture summaries. It offers a place to revisit concepts and work through material at a comfortable pace.

Deep learningInteractive notesJavaScriptHTML and CSS

Live guide

How I work

What matters to me when I'm building.

Understand the problem

Start with a conversation.

Useful feedback can change the starting idea. Conversations with professionals helped us refine DevTwin, while stakeholder discussions at Accenture guided automation choices.

User feedbackRequirementsPrototyping

Make choices

Choose an approach that fits.

For Second Brain, the challenge was keeping information useful without letting model costs grow unnecessarily. That led to different models for different tasks and updates only when content changed.

Model selectionCost and qualityArchitecture

Check the result

Make room for evaluation.

Building something is only part of the work. In research, we use comparisons and human feedback to understand the strengths and limitations of an approach.

Human evaluationStatistical analysisUser benefit

Work with the team

Keep the next step clear.

A useful plan makes priorities, responsibilities, and deadlines easy to follow. Whether coordinating delivery or workshop preparations, I try to keep those details clear and people informed.

PrioritizationTeam leadershipCoordination

Skills

Skills I use in my work.

Product discovery and delivery

Customer interviews, surveys, personas, and competitive research
Product requirements, roadmap priorities, KPIs, and Figma prototypes
Pricing, market sizing, and planning a product launch

DevTwin, Career Launchpad, GridSense, and stakeholder work at Accenture.

See related work

AI decisions and evaluation

Model selection, prompt engineering, and LLM benchmarking
Human evaluation, privacy controls, and output checks
Comparing quality, response time, and cost

Second Brain model choices, YOKE privacy checks, MedicalQA evaluation, and Beyond Flesch comparisons.

See related work

LLM systems, ML and NLP

RAG, embeddings, ChromaDB, BGE models, agents, and MCP tools
LoRA, PyTorch, and Hugging Face Transformers
Scikit learn, XGBoost, and Whisper

Research experiments, Career Launchpad retrieval, and Agent Shield Audio.

See related work

Applications and integrations

Python, TypeScript, Java, SQL, and Node.js
FastAPI, NestJS, React, REST APIs, GraphQL, and OAuth
PostgreSQL, Prisma, and Supabase

Blaze accounting integrations, YOKE application features, and the MedicalQA evaluation interface.

See related work

Cloud, automation and testing

Cloud Run, Vertex AI, Docker, GitHub Actions, and Jenkins
Jest, Selenium, Cucumber, Behave, RPA, and regression testing
Cloud cost reporting across AWS, Azure, and GCP

Enterprise automation at Accenture and cloud workflows at Blaze.

See related work

Analysis and coordination

Python, pandas, NumPy, SQL, and Power BI
Jira, Confluence, Google Sheets, and Agile ways of working
Stakeholder communication, dependencies, and team priorities

Research analysis, team delivery, and NextGen AI4Medicine planning.

See related work

Contact

Let's talk.

I'm interested in AI product management opportunities, including APM and RPM programs, and I'm also open to AI engineering roles. If my background fits your team's work, I'd be glad to connect.

Email me LinkedIn GitHub

Download my resumes

AI PM · PDF AI Engineer · PDF

The product resume covers AI PM, APM and RPM roles.