Available now · India + Germany
Apurv Adarsh
Product Manager — Internal Tools · Ops Automation · GenAI
Ex-Amazon PM intern (Munich, Pan-EU retail). Ex-engineer (DXC, 4 years). HHL Leipzig MBA.
I build analytics pipelines, workflow systems, and GenAI-assisted products that ops teams actually adopt —
and I measure adoption, not just delivery.
+30%
Forecast accuracy
Overstock KPI system
35→5 min
Banner build time
GenAI RAG tool
4 → 30+
Stakeholder adoption
Dashboard rollout
~30 min
Saved per user/day
Receipt automation
Case studies
4 projects
Global E-commerce Co.
Analytics · KPI Design
Overstock KPI & Dashboard System
Teams couldn't align on overstock drivers — decisions were slow, markdown risk was high, and the existing reports arrived too late to act on. I reframed this as a decision-quality problem, not a reporting one.
↓
+30%forecast accuracy
−10%markdown exposure
4 → 30+stakeholders adopted
What I owned
KPI tree definition, composite metric logic (stock health + weeks of cover + PO pipeline + demand trend), dashboard requirements, metric glossary, adoption plan. Partnered with BI throughout.
What I did
- Built an Overstock Probability metric — a composite signal that caught risk earlier than any single indicator alone
- Beta-tested with senior VMs for weeks before launch, iterating until prediction accuracy hit ~75%
- Embedded into existing WBR workflow (no new tool, no new login)
- Designed automated weekly digest — stakeholders arrived to reviews already informed
- Scaled adoption team-by-team: monitors team first, then PC super-team, then 30+ stakeholders
Tradeoffs managed
Metric trust risk → wrote public glossary with every input and threshold defined
Adoption risk → embedded in existing workflow, decision-view not just data-view
Data dependency risk → monitoring checks + explicit freshness header in digest
Composite vs simple → composite was necessary; WoC alone missed PO pipeline risk
Global E-commerce Co.
GenAI · RAG · Product Build
GenAI / RAG Deal-Banner Automation
Banner creation was slow, inconsistent, and expensive at scale. QA outcomes varied across brands with no systematic control. I owned the full PRD, model tradeoffs, retrieval design, guardrails, and phased rollout.
↓
35→5 minbuild time per asset
−40–50%cost per banner
80–85%QA pass rate (from 65–70%)
3 → 30+users scaled
What I owned
PRD + workflow design, model tradeoff decisions (quality / latency / cost), retrieval grounding strategy, guardrails design, eval criteria, rollout plan across brands.
What I did
- Baselined time, cost, QA pass rate, and adoption before building anything
- Designed GenAI + retrieval workflow — grounding outputs in approved templates and brand rules
- Defined eval criteria aligned to QA, iterated on quality checks with users
- Phased rollout: 1 brand pilot → 5 brands → 30+ users
- Fixed a data-leak issue with the security team mid-rollout without pausing adoption
Tradeoffs managed
Quality vs latency vs cost → retrieval grounding + hard constraints on model outputs
Hallucination risk → guardrails + template-pinning + human review step retained
Adoption trust → "options not answers" UX framing + QA-aligned acceptance criteria
Privacy/compliance → sanitised inputs/outputs, limited data exposure surface
Global E-commerce Co.
Workflow Automation · Ops
Email / Receipt Ingestion → Daily Digest
Manual processing of inbound emails and receipts was error-prone, time-consuming, and had no audit trail. I owned the pipeline design, failure-mode handling, and the "rules-based over AI" product decision.
↓
~30 minsaved per user/day
3 → 30–40users scaled
What I owned
Workflow schema (ingestion → extraction → validation → routing), failure-mode handling, daily digest format design, adoption plan.
What I did
- Designed the full ingestion pipeline with explicit failure modes and fallback paths
- Automated end-of-day digest format matched to existing team operating rhythm
- Deliberately chose rules-based over AI — prototype testing showed equivalent accuracy at lower cost and complexity
- Iterated on exception patterns and user feedback post-launch
Tradeoffs managed
AI vs rules-based → chose rules: equivalent accuracy, faster build, lower cost, easier to debug
Parsing edge cases → validation layer + manual review fallback for low-confidence extractions
Reliability → audit logs + retry patterns built in from the start
Adoption → digest format mapped exactly to how teams already closed their day
DXC Technology
Enterprise · Agile Delivery
Acting Product Owner — €2M Modernisation Roadmap
A legacy billing system modernisation for a national broadband programme was stalled by competing stakeholder priorities and painful deployments. I stepped into the PO role and rebuilt delivery predictability.
↓
50+requirements consolidated
+30%sprint velocity
−40%deploy time
−30%system errors
What I owned
Requirement alignment across 50+ stakeholders (3 business units), backlog prioritisation, sprint inputs, UAT/release readiness gates. Java + Camunda + ActiveMQ architecture context.
What I did
- Consolidated 50+ requirements into epics and a phased €2M roadmap
- Rebuilt refinement and acceptance criteria processes to increase sprint predictability
- Drove UAT and release gates to eliminate late-stage surprises
- Partnered on CI/CD improvements and cross-vendor integration delivery
- Restored project credibility after a team-lead gap — delivery stabilised within 3 months
Tradeoffs managed
Scope pressure vs delivery → phased roadmap with explicit tradeoffs negotiated with BT CIO
Speed vs quality → acceptance criteria + UAT gates as non-negotiable release conditions
Dependency risk → early alignment across BT, QA vendor, and offshore delivery teams
Active research — PR OS
India-first B2B SaaS · Ongoing
Product Discovery
India · B2B · Workflow OS
Active
PR OS — India-First PR Workflow System
A workflow system of record for India PR agencies — journalist CRM, follow-up queue, interaction timeline, and manager dashboard. The real pain isn't content creation; it's fragmented follow-ups across email, phone, and WhatsApp with no shared system of record. The wedge is operational discipline, not AI writing.
Key discovery findings
~30 journalists/week across 3–4 campaigns; reply rates 0–10 of every 100 outreach attempts
Follow-ups tracked in memory and email drafts — most journalists need 3+ touchpoints before responding
Channel sequence: email → phone → WhatsApp. No tool is built for this three-channel India reality
Managers have zero live view of ownership or campaign stage without calling the executive directly
Weekly reports rebuilt manually from email threads every Friday — takes 1–2 hours per person
14-day
pilot design ready
₹1k/mo
indicated WTP per user
MVP scope
Journalist CRM
Interaction timeline
Daily follow-up queue
Fast call logging
Templates
Campaign view
Manager dashboard
Reporting export
GitHub builds
github.com/apurv912
Skills & tools
Product
Discovery & Research
PRDs & User Stories
JTBD
Personas
RICE Prioritisation
Roadmapping
A/B Testing
Go-to-Market
AI / GenAI
RAG Pipelines
Prompt Engineering
Model Evaluation
Guardrails
LangChain
Amazon Bedrock
n8n
Analytics
SQL
KPI Design
Funnel Analysis
Dashboarding
GA4
QuickSight
Mixpanel
Delivery & Tools
Agile / Scrum
Stakeholder Mgmt
Backlog Governance
Jira
Figma
Confluence
Postman