Astra AI — Enterprise Revenue Intelligence — Full Architecture Three diagrams: main workflow, how RAG works in revenue context, how the LLM reasons over retrieved data Astra AI — Enterprise Revenue Intelligence End-to-end workflow Pipeline CSVs CRM deal data P&L Reports Financial PDFs Natural Language User query PyPDF + openpyxl Parse & chunk docs Gemini Embeddings Vectorise content LangGraph Orchestrate query flow Supabase pgvector Store & retrieve vectors Groq LLaMA 3 Generate answer LangGraph Router AE vs CFO persona split AE View Risk score + next action CFO View P&L + NRR answer Plotly Charts & visuals How RAG works — Revenue context From raw revenue data to semantically retrieved chunks Offline ingestion Pipeline CSV Deals, stages, ARR P&L PDF Revenue, margin, NRR Chunker Split into 512-token segments chunk_01: "APAC ARR $4.2M, down 12%…" chunk_02: "Deal ID 882, stage: negotiation…" chunk_03: "Gross margin Q3: 61%, vs 68%…" Gemini Embeddings 768-dim vector per chunk Supabase pgvector Indexed vector store Real-time retrieval "Why is APAC margin down?" Natural language query Gemini Embeddings Embed query → vector Cosine similarity search Match query vec vs stored vecs Return top-k chunks ✓ chunk_03: "Gross margin Q3: 61%…" ✓ chunk_01: "APAC ARR $4.2M, down 12%…" chunk_07: "US pipeline up 8%…" (low) Pass to LLM with context Prompt = query + top-k chunks vector lookup How the LLM reasons — Revenue intelligence From retrieved chunks to grounded, cited, persona-aware answer Prompt sent to Groq LLaMA 3 System prompt "You are a CFO assistant. Answer using ONLY the context provided. Cite source chunk IDs." Retrieved context [chunk_03] Gross margin Q3: 61%, vs 68% prior yr [chunk_01] APAC ARR $4.2M, down 12% QoQ User question "Why is APAC margin down this quarter?" Groq LLaMA 3 — Reasoning Cross-reference chunks · infer causality LangGraph Persona Filter Shape answer for AE or CFO lens AE answer "Deal 882 at risk — APAC ARR down 12%. Recommend exec sponsor call. [src: chunk_01]" Next-best action CFO answer "Gross margin declined 7 pts YoY driven by APAC mix shift. Reforecast NRR. [src: chunk_03]" Source cited Hallucination guard LLM restricted to retrieved context only — no invention Astra AI — Revenue Intelligence · Full Architecture Suite
Astra AI — Enterprise SDR Agent — Full Architecture Three diagrams: main SDR agent workflow, how RAG works for ICP and company research, how the LLM writes and scores personalised outreach Astra AI — Enterprise SDR Agent End-to-end workflow Lead List Input Company + contact targets LangGraph — ReAct agent loop Tavily Search Research company live Groq LLaMA 3 Score ICP fit Groq LLaMA 3 Write personalised email LangGraph State Manager Sequence & track follow-ups Human-in-the-Loop SDR approves before send Rejected Approved Revise draft Loop back to agent Email Sent Outreach delivered Supabase Log decisions & history How RAG works — SDR company research From internal playbooks + live web data to ICP-matched context Offline ingestion — sales knowledge base ICP Profiles Ideal customer criteria Sales Playbook Personas, pain points Chunker Split into 512-token segments chunk_01: "ICP: SaaS, 200–2000 employees…" chunk_02: "Pain: manual outreach, no signal…" chunk_03: "Persona: VP Sales, RevOps lead…" Gemini Embeddings 768-dim vector per chunk Supabase pgvector Internal knowledge store Real-time retrieval — per lead Target: "Acme Corp" Lead from input list Tavily Search Fetch live company signals → "Acme raised Series B, $40M" → "Hiring VP Sales + 3 AEs" → "Using Salesforce, Outreach" Cosine similarity match Acme signals vs ICP chunks ✓ chunk_01: ICP match — SaaS 800 emp ✓ chunk_02: Pain signal confirmed vector lookup Merged context → LLM ICP chunks + live signals combined How the LLM works — SDR personalisation engine From merged context to ICP score, personalised email draft, and HITL decision Prompt sent to Groq LLaMA 3 System prompt "You are an SDR agent. Score ICP fit 1–10 then write a personalised cold email. Use only provided signals." Retrieved context [chunk_01] ICP: SaaS, 200–2000 employees [chunk_02] Pain: manual outreach, no signals Live signals (Tavily) Acme: Series B, $40M Hiring VP Sales + 3 AEs Stack: Salesforce + Outreach Groq LLaMA 3 — Two tasks Score ICP fit · Draft email copy ICP score: 8.5 / 10 + SaaS, 800 emp ✓ matches ICP + Hiring AEs = revenue growth + Outreach in stack = pain fit High priority Email draft "Hi [Name], saw Acme just closed Series B — congrats. As you scale the sales team, we help teams like yours…" HITL Review SDR edits, approves, or rejects Personalised outreach sent Supabase logs score + draft + outcome
Astra AI — Multi-Agent Command Centre Architecture Complex query routed by LangGraph orchestrator to specialist CrewAI agents then synthesised into one response Astra AI — Multi-Agent Command Centre Complex Enterprise Query Multi-domain question input LangGraph Orchestrator Decompose task, route to agents Sales Agent Pipeline & revenue HR Agent Policy & people data Finance Agent P&L & forecasts Research Agent Live web lookup Tavily Search Real-time web data Groq LLaMA 3 Fast parallel inference CrewAI — specialist agents run in parallel LangGraph Synthesiser Merge & compose final answer Executive Response + Audit Trail Supabase logs every agent step
Astra AI — AskHR Policy Assistant Architecture HR PDF ingested via PyPDF, embedded with Gemini, stored in Supabase pgvector, retrieved by LangChain RAG, answered by Groq with source citation Astra AI — AskHR Policy Assistant Ingestion pipeline (one-time) Query pipeline (real-time) HR Policy PDFs Leave, benefits, 401k docs PyPDF Parse & chunk PDF pages Gemini Embeddings Vectorise text chunks Supabase pgvector Store vectors + metadata Employee Question "How many leave days do I get?" Gemini Embeddings Embed query for retrieval semantic match LangChain RAG Chain Retrieve top-k chunks Groq LLaMA 3 Generate cited answer Policy Answer + Source Citation Exact document & section referenced
Astra AI — Evals and Observability Architecture RAG response scored by RAGAS for faithfulness and relevancy, traced by Langfuse and LangSmith, drift monitored by Evidently AI Astra AI — Evals & Observability User Query Question to RAG system RAG Pipeline Retrieve → Groq LLaMA 3 → Generate LLM Response Answer + retrieved context RAGAS Faithfulness score Relevancy · Context recall Langfuse · LangSmith Trace every LLM call Latency · tokens · cost Evidently AI Drift detection Model quality over time Groq LLaMA 3 LLM-as-judge hallucination Live Eval Dashboard Scores · traces · drift alerts in Streamlit
Astra AI — Copilot and Insights Architecture User query routed by LangChain ReAct agent to search, math or content tools, response returned with live observability metrics Astra AI — Copilot & Insights User Query Search · maths · email request LangChain ReAct Agent Reason → decide tool → act Tavily Search Live web search Python Tool Execute maths / logic Gemini Flash Draft email / content Groq LLaMA 3.3 70B Final response generation tiktoken Count tokens exactly Copilot Response Answer shown in chat UI Supabase persists history Live Insights Panel Latency · tokens · cost Confidence · tool used · success rate
Lead Product Manager — AI / GenAI Platforms

Enterprise AI Strategy.
Disciplined execution.
Measurable impact.

11+ years building Enterprise SaaS platforms for APAC Fortune 500 clients. I bridge AI strategy and execution, defining which AI opportunities matter the most across RAG, AI copilots, multi-agent systems, evaluation, and responsible AI.

Astra — Live AI Product Demos
Karamjeet Singh
$100M+AI-Optimized — Nestlé India Portfolio
3 days → 20 minSpend Insight GenAI Co-pilot
₹32,705 CrAUM — Bajaj Housing (0→1)

Astra — Live AI Product Demos

Real applications, running in production, built end-to-end.

Persona-aware RAGLangGraph Orchestration
●Live
Astra AI — Enterprise Revenue Intelligence

CFOs lack live P&L intelligence, while AEs lack real-time deal coaching. Astra AI enables both to ask natural-language questions like “Why is APAC margin down?” and receive instant, cited, data-grounded answers with risk scoring and next-best-action recommendations — without analysts in the loop.

TECH STACK:
LangGraphGroq LLaMA 3Gemini EmbeddingsSupabase pgvectorPlotlyPyPDFStreamlit
▶ Launch DemoGitHub →
Agentic WorkflowHuman-in-the-LoopReAct Agent
●Live
Astra AI — Enterprise SDR Agent

Sales reps lose valuable selling time to research and outreach preparation. Astra autonomous AI agent identifies leads, scores ICP fit, researches accounts, drafts personalized emails, and orchestrates follow-up sequences with human-in-the-loop approval.

TECH STACK:
Apollo.ioTavily SearchLangGraphGroq LLaMA 3StreamlitPython
▶ Launch DemoGitHub →
Multi-Agent OrchestrationAgent-to-Agent
●Live
Astra AI — Multi-Agent Command Centre

Master orchestrator that interprets complex enterprise requests, decomposes them into sub-tasks, routes work to specialist domain agents (Sales · HR · Finance · Research), and synthesizes a single executive-ready response with a complete audit trail.

TECH STACK:
LangGraphCrewAIGroqGeminiSupabase
▶ Launch DemoGitHub →
RAG PipelineSemantic SearchLLM Eval
●Live
Astra AI — AskHR Policy Assistant

Employees waste hours searching policy docs. Astra RAG-powered assistant answers any policy question by semantically searching actual HR documents — not generic knowledge — and cites the exact source.

TECH STACK:
Gemini EmbeddingsSupabase pgvectorLangChainGroqPyPDF
▶ Launch DemoGitHub →
RAGAS EvalHallucination Detection
●Live
Astra AI — Evals & Observability

Enterprises can’t trust AI answers. Astra live evaluation dashboard running RAGAS metrics on every RAG response, plus a hallucination demo with red-teaming — shows AI failing and why, building trust through transparency.

TECH STACK:
RAGASLangfuseLangSmithGroqGemini
▶ Launch DemoGitHub →
Live AI ObservabilityCost & Latency
●Live
Astra AI — Copilot & Insights

LLM-powered copilot unifying search, calculations, and content generation, with a live observability layer showing latency, token usage, hallucination risk, and cost metrics per query and across the session.

TECH STACK:
Groq LLaMA 3.3 70BGemini FlashTavilySupabasetiktoken
▶ Launch DemoGitHub →

AI Product Approach

The approach behind every module in Astra AI and every AI system I’ve shipped is treating trust as a core product layer.

01
Evals before prompts

I define what “good” looks like before writing the first system prompt. Hallucination rate, retrieval precision, response latency — all scoped upfront as product requirements, not afterthoughts.

02
HITL as a first-class interface

Human-in-the-loop isn’t a disclaimer. It’s a screen I design, a decision I log, and a feedback signal I feed back into the model.

03
Explainability over accuracy

A system that’s 95% accurate but can’t explain its reasoning will get abandoned. A system that’s 88% accurate and shows its sources will get trusted and iterated on.

04
Architecture is a PM decision

Choosing RAG over fine-tuning, multi-agent over single LLM, rules-based scoring over pure ML — these are product calls with tradeoffs in cost, latency, maintainability, and trust. I own them.

Case Studies

Deep dives into work highlights — tap to expand for full context, approach, and learnings.

01
Nestlé India MarTech AI Transformation
Multi-Agent AI · GenAI Co-pilot · $100M+ Portfolio
70% AI Adoption45% Faster Planning3 Days → 20 Mins
▾
Context

Nestlé India's $100M+ MarTech portfolio bottlenecked by manual workflows — CMOs waited 3 days for spend insights, AI adoption near zero across brands.

My Role

Lead PM across 4 squads, 30+ members. Owned AI strategy, roadmap, delivery across Digi-Cadence, Maxx-RoI, and Meta-360. Accountable to Nestlé & Unilever stakeholders.

Challenge

Driving AI adoption at Fortune 500 scale with hallucination risk, data governance concerns, and need for interpretable AI CMOs would actually trust.

Approach
  • Production RAG co-pilot over 150+ brand Cadence Scores — <4% hallucination
  • Multi-agent workflow (SOV, CPC, MMM) with HITL override
  • AI evals & governance — automated regression and red-teaming
  • Shared model serving layer — 3× faster AI feature velocity
Result
  • 70% AI adoption across Nestlé India's MarTech portfolio
  • 45% faster planning, time-to-insight 3 days → 20 mins
  • ROAS acceptance +35%, forecast accuracy +30 pts
  • 92% client AI trust score
Key Learning

"Enterprise AI adoption lives or dies on trust. Eval frameworks and HITL design aren't guardrails — they're the product."

02
Bajaj Housing — 0→1 Enterprise Lending Platform
Founding PM · LOS/LMS · Origination to Disbursement
₹32,705 Cr AUM100K+ Customers350+ Partners
▾
Context

Bajaj Housing Finance needed an enterprise lending platform from scratch — full loan lifecycle from origination through underwriting to disbursement.

My Role

Founding PM. Owned product strategy and delivery for LOS/LMS with credit, risk, legal, and operations teams.

Challenge

Regulated financial platform at scale — zero legacy, complex underwriting, hard go-live with regulatory compliance requirements.

Approach
  • Full LOS/LMS architecture — origination, underwriting, disbursement
  • Industry-first pro-rata eligibility engine — borrower base +30%
  • Onboarded 350+ builder partners into ecosystem
Result
  • ₹32,705 Cr AUM within 2 years, 100K+ customers
  • 350+ builder partner integrations
Key Learning

"The eligibility engine wasn't a feature — it was a business model unlock hiding inside a risk framework."

03
VTS Rise — Empire State Building Amenity Platform
PropTech · 1,500+ Buildings · 3M+ Occupants · 800K+ MAU
800K+ MAUEmpire State Building3M+ Occupants
▾
Context

VTS needed to launch ESRT+ at the Empire State Building — serving enterprise tenants across 65,000 sq ft of premium workspace.

My Role

Senior PM for VTS Rise. Owned amenity booking, billing infrastructure, bluetooth access, real-time tracking across 1,500+ buildings.

Challenge

Consumer-grade UX for enterprise real estate — multi-currency billing, Apple Pay, real-time space management at global scale.

Approach
  • Amenity booking with bluetooth mobile access
  • VTS Activate billing — Apple Pay, multi-currency, bulk invoicing
  • Real-time occupancy tracking across 1,500+ buildings
Result
  • 800K+ MAU, 3M+ occupants, Empire State Building 65,000 ft²
Key Learning

"The tension between consumer simplicity and enterprise compliance is where real PM work lives."

04
Smartworks — Pan-India Occupancy Intelligence
PropTech · ML Pricing · 50+ Centres · 203K+ Seats
203K+ SeatsUtilisation +18%Revenue +25%
▾
Context

Smartworks operated 50+ centres with 203K+ seats — manual pricing led to idle inventory and unpredictable revenue.

My Role

Senior PM. Owned occupancy tracker, pricing engine, and revenue forecasting products.

Challenge

Real-time visibility across 50+ dispersed centres — replacing gut-feel pricing with ML ops teams would trust.

Approach
  • Pan-India Occupancy Tracker with centralised ops dashboard
  • ML-driven dynamic pricing via NLP pipeline
  • Revenue forecasting model
Result
  • Utilisation +18%, idle inventory −15%, revenue predictability +25%
Key Learning

"Making the model’s logic visible and overridable was the unlock for adoption."

Career Timeline

11+ years building enterprise products across FinTech, MarTech, and PropTech.

2023 – Present
KiE Square / Nestlé
Lead PM · AI/GenAI
2022 – 2023
Smartworks
Senior PM
2021 – 2022
BigStep / VTS
Senior PM
2020 – 2021
KiE Square / Unilever
Senior PM
2018 – 2020
Bajaj Housing
Product Manager
2015 – 2018
HDFC Credila
Associate PM
KiE Square Analytics  /  Nestlé
May 2023 – Present
Lead Product Manager · 4 Squads · 30+ Members
  • Multi-agent AI workflow — ROAS +30 pts, analyst hours −40%
  • GenAI Co-pilot (RAG) — 3 days → 20 mins, <4% hallucination rate
  • AI evals & governance — 92% client AI trust score
  • KiEVerse AI platform roadmap — 4 squads, 3× faster feature velocity
Smartworks
Jul 2022 – May 2023
Senior Product Manager · Pan-India Coworking Platform
  • Pan-India Occupancy Tracker — 50+ centres, 203K+ seats
  • ML-driven pricing — revenue predictability +25%, idle inventory −15%
BigStep Technologies  /  VTS
May 2021 – Jun 2022
Senior Product Manager · PropTech · 1,500+ Buildings
  • ESRT+ at Empire State Building — 65,000 ft² tenant experience platform
  • VTS Activate billing — Apple Pay, multi-currency; 800K+ MAU
KiE Square Analytics  /  Unilever
May 2020 – May 2021
Senior Product Manager · Unilever Onboarding
  • Onboarded Unilever onto Digi-Cadence — feature adoption +30%
  • Marketplace integrations — Flipkart, Nykaa, BigBasket
Bajaj Housing Finance
May 2018 – May 2020
Product Manager · Founding PM · LOS/LMS Platform
  • 0→1 lending platform — ₹32,705 Cr AUM, 100K+ customers
  • Pro-rata eligibility engine — borrower base +30%
HDFC Credila
Jun 2015 – May 2018
Associate Product Manager · FinTech · Education Loans
  • Digitized education loan journey — KYC automation, verification −20%
  • SQL dashboards for lead funnel & conversions

About

I build AI products that solve real business problems.

With 11+ years leading product strategy across FinTech, MarTech, PropTech, and enterprise SaaS, I've built AI-driven products, revenue platforms, and workflow automation systems that drive measurable outcomes. I focus on identifying what matters and why, balancing technical possibilities with business impact, and building products teams actually adopt.

My developer background allows me to prototype AI solutions hands-on, moving quickly from idea to working product while thinking in systems and tradeoffs. I also build AI products independently to experiment with emerging technologies, validating what works in practice before applying those learnings to enterprise strategy.

I start with the problem first, looking for business challenges that represent defensible ROI opportunities for AI investment. I believe in establishing decision frameworks before scaling, building with production realities in mind from day one, and designing fallback strategies for when systems fail. What I care about: business metrics over technical metrics, unit economics over feature count, and sustained impact over launch hype.

Education

Academic foundation across business strategy and technology.

MBA
Marketing & IT
Balaji Institute of Modern Management, Pune
B.Tech
Information Technology
Guru Gobind Singh Indraprastha University, New Delhi

AI Tech Stack

GenAI engineering layer — from model selection to production observability.

LLMs
GPT-4oClaude SonnetGemini 1.5 FlashLlama 3.3 70BGroq
Infra
Azure OpenAIAWS BedrockAnthropic APIHuggingFaceVS CodePoSH
RAG & Vector
LangChainLlamaIndexPineconeSupabase pgvectorTavilyPyPDFGemini Embeddings
Agents & Orchestration
LangGraphCrewAIn8nReActApollo.ioHITL Design
Evals & Observability
RAGASPromptfooEvidently AILangfuseLangSmithDrift DetectionRed-teamingtiktoken
Visualisation
StreamlitPlotlyTableauPower BIMixpanel
Product & Strategy
AI Platform StrategyModel SelectionGuardrailsResponsible AIAI Product RoadmapMLOps
Leadership & Data
OKRsA/B TestingStakeholder ManagementCross-functional LeadershipSQL

Let’s Connect.

Open to AI Product Manager roles — happy to discuss AI strategy, product execution, and building AI-powered products that drive real outcomes. Remote, Hybrid, or On-site.

karamjeet.6613@gmail.com Resume New Delhi, India