I turn messy, rule-heavy workflows into AI products people trust.
9+ years across payroll, fintech, insurance, edtech and construction. I've shipped custom document AI at Sage, agentic AI and MCP at Zywave, and three products of my own. Scroll to play through it.
AI Product ManagerFounder ×3EducatorSpeakerExplorer
9+years in product
700+agencies on my agent
96%doc AI accuracy
3×Best PM of the Year
Scene 1 of 5 · Loading player
Scroll ▼
Character select
Your player: AI product leader with founder scars
Each ability is backed by something I shipped. The levels below show the proof.
P1Class: Builder
GV SAGARPune, India · MCA, VIT (AI/ML)
Product strategy0-to-1 roadmaps, TAM/SAM/SOM, AI economics
AI & LLM systemsCustom models, RAG, MCP, prompt and caching architecture, evals
GTM & monetizationToken and credit-based pricing, release process across Sales, Marketing and Engineering
Team buildingHired the India product team, coached POs, Agile transformation (10× delivery)
Teaching & stageLearnEzzy AI cohorts, industry expos, AI training for tech and non-tech teams
Special moves
Model Context ProtocolRAG pipelinesPrompt + caching architectureAI cost vs value metricsHuman-in-the-loop designData fabricSAFe POPMFigma · Jira · Confluence
World map
Four worlds, one bonus stage
As you scroll, my character travels the path. Enter any level to play through the mission, the boss fights, the power-ups and the loot.
1
World 1 · 2018–2020Enterprise payroll
The rules engine behind every paycheck
Deloitte · Sr. Analyst · Deloitte Time & Expense
Timesheets, expenses, payroll and client billing for a workforce spread across many projects and clients at once, with a different holiday rulebook in every US state.
A custom, in-house extraction model for invoices and bank statements, built before generative AI existed and ramped from 10% to 80% of daily documents.
Designed from scratch: research, UX, prompt design, tool and skill evals, and the teams behind them. Now the agent works inside Claude, ChatGPT and Copilot.
Deloitte Time & Expense (DTE) decides how every employee gets paid and how every client gets billed. My job was to make the product's rules smarter without breaking a payroll run.
Existing product: my work here was enhancements and new features
1
01 · Mission
One timesheet, five systems downstream
Deloitte runs many projects for many client companies at the same time. Each week every employee fills a timesheet: which projects they worked on and how many hours went to each.
Those hours drive expenses, then employee pay, then client billing, including customers on annual maintenance contracts (AMCs). The same data flows to HR for payroll and org management. Without DTE, Deloitte can't run compensation.
INWeekly timesheetProject × day × hours
RULESState policy engineHolidays, leave, overtime by state
OUT 1ExpensesCalculated from time
OUT 2Payroll & HRCompensation, org data
OUT 3Client billingIncluding AMC customers
02 · Play the system
Switch states and watch the money move
This is a simplified model I built for this page (DTE's real interface isn't public). Thursday is a holiday. Switch the employee's state, or log hours on the holiday, and see how payroll, billing and expenses respond.
DTE · week 47 · illustrative
Project
Mon
Tue
Wed
Thu · holiday
Fri
Total
Payroll · to HR0 h
Client billing$0
Expense claim$0
States A, B and C and their rules are made up to show how the logic works. They don't describe real state law or Deloitte policy.
03 · Boss fights
Where the rules fought back
BOSS 1
The rulebook keeps changing
When a state updated its holiday or leave policy, hard-coded logic broke timesheets, expense calculations and invoices all at once.
BOSS 2
Same week, different math
Two consultants on one client project take the same holiday. They live in different states, so their pay and billable hours come out differently, and both have to be right.
BOSS 3
Errors show up too late
A wrong entry stayed invisible until payroll or invoice reconciliation, weeks later, when fixing it took several teams.
04 · Power-ups
What we shipped
POWER-UP 1
State-aware rules at the point of entry
Each state's holiday and leave rules apply while the employee fills the timesheet, not after submission.
POWER-UP 2
Validation before submit
Hours on a holiday, missing project codes and over-limit days get flagged while the employee can still fix them.
POWER-UP 3
Guidance in context
Clear instructions next to each holiday type, so employees know how to log a holiday on a billable project.
POWER-UP 4
One source of truth downstream
Payroll, HR and client billing read the same calculated numbers, so reconciliation stops being a monthly battle.
How I worked: I turned the product vision into user stories and backlog items developers could build without guessing, prioritized them by ROI with Product Management, checked technical feasibility with Product Owners, and worked with support to diagnose production issues and decide which to fix first.
05 · Loot
What changed
Per state
Rules applied where the employee makes the decision
Earlier
Errors caught before submission instead of at payroll
1 truth
Payroll, HR and billing working from the same numbers
What this level taught meIn a rule-heavy system, the rule is the product. Put it where the user makes the decision, not where the accountant finds the mistake.
I managed AutoEntry, Sage's AI accounting automation product, working with the core backend team and the AI teams. Customers upload invoices, receipts and bank statements; AutoEntry extracts the data and publishes it to their accounting software.
Built 2021–2023, before generative AI: every component was designed and tested by hand
2
01 · Mission
Replace manual extraction without risking financial data
Users upload invoices, sales, expenses and bank statements. The documents went to an internal system where data processors extracted the fields, mostly by hand and some with OCR.
This is financial data, so sending it to an external AI model was off the table. We had to build our own custom model. That meant this wasn't a wrapper: we also had to plan custom deployment and infrastructure.
Implementing AI sounds easy. The real work is knowing where AI belongs and why. So we started with a gap analysis, not a model.
02 · Gap analysis
Every upload was one of three shapes
SHAPE 1Single page, single invoiceReady for extraction as is
SHAPE 2Multi-page, multi-invoiceNeeds splitting and merging into the right invoices
SHAPE 3Single page, multi-invoiceNeeds cropping into separate invoices
Before processing, we have to capture the primary fields: invoice number, supplier name, dates, totals, tax and line items. When one invoice contains several images, both the customer and the AI get confused. Documents had to be trimmed, cropped and merged first.
03 · Play the pipeline
You're the data processor. Feed it a document.
This is the pipeline we designed. A router model decides whether a document needs fixing. If it does, the AI crops or merges it and a human approves the result. Then the custom model extracts the fields. Sample documents are fictional.
AutoEntry pipeline · illustrative
STEP 1UploadWeb, mobile or email
STEP 2Router modelNeeds merge, crop or split?
STEP 3AI fixer + human reviewApprovals measure accuracy
Several invoices in one image, or one invoice spread across pages. Data gaps and missing information on top of that.
BOSS 2
No generative AI to lean on
In 2021–2023 every component was designed, labelled and tested by hand. That's why a safe ramp took six months.
BOSS 3
Hallucination at scale
When we increased throughput, the model started inventing values. On financial data, one invented total is a broken book.
05 · The ramp
From 10% to 80% of daily documents in six months
We increased the load in steps. When hallucinations appeared, we stopped, fixed them, then grew the load slowly again while a feedback model improved accuracy.
Share of daily documents through AIModel accuracy
The start and end values (10%, 80%, 96%) are real. The path between them shows the shape of the journey, not exact monthly figures.
06 · Power-ups
What we built
POWER-UP 1
Router model
Identifies which documents need merging, cropping or splitting, so the extractor only sees clean, single invoices.
POWER-UP 2
AI fixer with human approval
The AI merges or crops; data processors approve. Every approval also scores the model.
POWER-UP 3
Accounting-native extraction model
A custom model that understands accounting documents: each one is converted to images, then the fields are extracted.
POWER-UP 4
Feedback model
Corrections flow back into training, which is how accuracy kept rising after the hallucination fix.
POWER-UP 5
Controlled ramp
Load grew from 10% only while accuracy held. Speed was never allowed to outrun trust.
POWER-UP 6
Team and process
I led AI and backend teams, ran an Agile transformation across India R&D and coached POs.
07 · Loot
What we won
96%
Model accuracyon 80% of daily inflow
80%
Daily documents handled by AIup from 10% in six months
<4 hrs
CA audit turnarounddown from 7–8 days with an AI invoice audit product for SMEs
200K+
Active usersserved by the products I worked on
10×
Delivery speedafter the Agile transformation
+23%
Firm revenueroadmap delivery contribution
What this level taught meKnowing where AI fits mattered more than the model. The router kept the model away from problems it was never built to solve.
I own Apex MCP and Producer Agent end to end. I designed them from scratch: the research, the UI and UX, how the prompts work, the tool and skill evals, and the teams that build them.
Ongoing: this level is still being played
3
01 · Mission
Give producers their time back
Insurance producers win business through relationships, but most of their week goes to research, list building and cold emails. Zywave says more than 60% of producer time goes to back-office work that technology can handle.
I spotted the shift early by combining customer behaviour data with competitive intelligence from several market-research tools. Then I pitched it and won leadership buy-in to build an AI-powered CRM, enrichment and outreach platform.
Apex MCPZywave's insurance data and actions, available inside Claude, ChatGPT, Copilot or any MCP-compatible AI.
Producer AgentFrom the ideal customer profile to a booked meeting: research, outreach, tracking and next steps.
Advisor AgentThe third part of the Apex platform, focused on the existing book of business.
02 · Play the agent
A producer asks Claude. The agent does the work.
A replay of how the workflow feels inside an AI assistant. Company names, data and tool names here are fictional.
AI assistant + Apex MCP · replay
The producer keeps final approval on every email.
03 · Boss fights
Five bosses on the way to production
BOSS 1
Trust
Producers won't let a machine email their prospects unchecked. Generic LLM output also gets insurance details wrong.
BOSS 2
Adoption
One more login means one more tool nobody opens. Agencies already work in Outlook, their AMS and their AI assistant.
BOSS 3
Cost
Token costs grow quickly at the scale of hundreds of agencies running sequences every week.
BOSS 4
Quality
How do you know a tool or skill still works after a prompt changes? Gut feel doesn't scale.
BOSS 5
Pricing
Large upfront AI pricing blocked customers from trying it at all.
04 · Power-ups
How each boss went down
VS TRUST
Human approval and grounded data
Every email waits for the producer's approval, and the content is grounded in Zywave's own data and expert-written material.
VS ADOPTION
The agent lives inside their AI
Through MCP, Producer Agent's full workflow runs inside Claude, ChatGPT and Copilot. No new login.
VS COST
Prompt and caching architecture
Designed with data science to cut compute cost, plus an org-wide view of AI usage against cost and benefit.
VS QUALITY
Tool and skill evals
Evals for every MCP tool and skill, and RAG and LLM experiments to prove feasibility before we build.
VS PRICING
Modular token and credit pricing
A shift from large upfront deals to pay-for-what-you-use AI features, which opens new upsell paths.
TEAM
The India product team
Hired 3–5 PMs and POs, and designed a release process that keeps Sales, Marketing and Engineering in step at launch.
05 · Loot
The scoreboard so far
3% → 8.2%
Cold-email KPIflagship deployment across 700+ insurance agencies
+3–9%
Conversion liftacross deployments
~6 hrs
Manual work saved per workflowtwo workflow automations, now top performers
30 → 1
Data fabric30 products unifying into one AI data suite (in progress)
I also built an AI-powered legal and finance quoting (BOQ) tool for the AECO domain, and I present our AI use cases at industry expos and conferences.
What this level taught mePut the agent where people already work, and keep their hand on the send button.
Outside my day job I build my own products. Each one started with a workflow I watched break: civil engineers learning from outdated material, construction firms running on Excel and WhatsApp, sales teams with stale lead lists.
4
Stage 4-1 · EdTech
CivilGuruji
CIVIL की बात, GURUJI के साथ
LMSCommunitySales & marketing pipeline2 lakh+ active users
The problem: Fresh civil engineers graduate knowing theory but not the site: estimation, quantity surveying, layouts, AutoCAD, BIM. Practical training was offline, expensive and limited to a few cities.
What I built: An LMS for practical, job-ready civil engineering courses, a community where learners ask and answer site questions, and a sales and marketing pipeline that takes an enquiry through counselling to enrolment.
Boss fight: learners study on their phones between site shifts, often on weak connections. Lessons had to be short, light and in the language they think in.
Rohit K. · site engineer What steel ratio do you use for a 230 mm brick wall lintel?
AP
Anjali P. · mentor Start from the span and the load. Here's the estimation sheet from Module 6.
Stage 4-2 · AECO automation
Shegruh
Run your whole design or construction business in seconds, with AI.
Lead managementSheAIAccountingMaterial managementSite reporting
The problem: I also run a construction business, so I've seen it first-hand. Quotes live in Excel, site updates in WhatsApp groups, purchase orders in notebooks. Interior designers, architects and construction firms lose days to paperwork.
What I built: One platform for leads, quotations and proposals, accounting and invoices, purchase orders and inventory, and daily site reports. SheAI writes project timelines, proposals and legal and contract documents, and generates a detailed BOQ in about two minutes.
Boss fight: small firms won't pay for features they don't use, so pricing is pay-as-you-go: pick features, users, AI credits and storage.
Loot: live for firms in India, the US, the UK, the UAE, Ireland, Japan and Singapore.
SheAI › 2BHK interior, 950 sq ft, modular kitchen, false ceiling in living room
Item
Qty
Unit
Amount ₹
Estimated total
₹0
Stage 4-3 · CRM + sales
Dugong
Find the right buyer, enrich them and write the first line, in one table.
Lead databaseWaterfall enrichmentAI research columnsSequencesPipeline CRM
The problem: sales teams buy one tool for leads, another for enrichment and a third for outreach. The data goes stale between them.
What I built: a CRM and sales app in the Apollo and Clay style. Every row is a prospect, every column is an enrichment step: firmographics, the decision maker, a verified email, an intent score and an AI-written first line.
Boss fight: no single data source is complete. A waterfall tries several providers in turn until a field is filled, and charges credits only when a cell succeeds.
Dugong · Table: Q1 targets
Company
Employees
Decision maker
Email WATERFALL
Intent AI
First line AI
What this world taught meFounders don't get to choose their problems. I build for the workflow I've seen break, then let the users tell me what I got wrong.
Bring your hardest problem: an AI career move, a sales pipeline that won't grow, a launch with no plan, a product nobody's sure fits. We work on it together, and you leave knowing your next three moves.
P2
01 · Choose your quest
What should we work on?
Pick up to two topics. Your choices are added to the booking form below.
02 · How the hour runs
Four rounds, one boss: your problem
BEFOREBriefYou fill in the form below so I arrive prepared.
0–10 MINContextYour goal, your constraints, what you've already tried.
10–45 MINDeep workWe work the problem: frameworks, examples, live teardown.
45–60 MINGame planYour next three moves, prioritised, written down together.
03 · Book
Request your session
Tell me a little about yourself and pick a preferred time. I'll reply by email to confirm the slot.
If your email app doesn't open, copy the request and send it to iamgvsagar@gmail.com, or message me on LinkedIn. Your request is only sent when you send that email.