Mridul Gupta

Fintech Architecture & Platform

• Own architecture and engineering direction for a multi-bank, multi-geography fintech platform operating across the US, Canada and UK — scalability, reliability, latency, cost and regulatory requirements.
• Data residency & governance by design — data stored and processed per operating-country requirements, with the platform flexible enough to add geographies without significant architectural rework.
• Payment routing engine optimizing transaction routing and settlement — lifted real-time settlement coverage from 10% to 90% in six months, replacing a cycle that took up to 3 days, while lowering payment-processing and collection costs.
• Lightweight CQRS layer in front of the highest-traffic services — app launch and first-load latency cut from 500ms to under 120ms without rearchitecting the platform.
• Realtime rule & signal-processing platform on Apache Flink for fraud analysis, realtime compliance and credit monitoring — the fraud pipeline already shows a +20% lift in anomaly detection; also worked with third-party fraud providers on realtime detection.
• Multi-country onboarding for customers moving across Canada, the UK and US — country-specific requirements vs onboarding cost vs approval rates.
• Feature store supporting ML models including churn prediction (78% recall, 0.83 F1, ~89% precision) and customer LTV — a common layer for managing and serving features.
• Leading the migration to Databricks from an internally built warehouse with drifting data quality — company-wide data access and the observability a fintech needs from its own data.

AI & Agentic Systems

• Built Zolve's AI capability from the ground up — starting with RAG-based systems and natural-language-to-SQL for business analytics.
• Semantic data layer with deterministic query generation — business users ask questions in natural language while schema and data stay under control.
• Internal no-code agent platform — single-task agents, agent-to-agent handoffs and multi-step agentic workflows, with session management, memory, observability and analytics.
• Scaled to 100+ production agents across dispute management, churn management, customer support and escalations — 25% of disputes resolved end-to-end and 90% of compliance reporting & QC now run on agents, with humans reviewing exceptions.
• AI voice collections — ElevenLabs voices orchestrated on the Retell AI agentic platform, outperforming human collectors by 72% in the lower bucket, the earliest and highest-volume stage of delinquency.
• Fraud-detection dashboards in Metabase on a governed data layer connected to Anthropic models to surface fraud patterns.
• Engineering layer on LangGraph and LangSmith — reliability, observability, cost management, sessions, memory and continuous improvement.

Product Velocity & Mobile

• Native iOS + Android → React Native — release timelines cut by 50% and time-to-ship for new features by 60%, where the banking product didn't need realtime-classroom-grade platform complexity.
• Engineers per release: 8 → 2–3 — roughly a 70% smaller footprint per ship once both platforms converged on one codebase.
• No-code release platform — product managers compose screens and features from predefined widgets and templates, decoupled from the traditional mobile-release cycle.

Engineering Leadership

• Work directly with the founder and leadership on technology investment, engineering economics, acquisition, retention, churn and P&L trade-offs; communicate technology strategy and business impact to investors.
• Build teams around individual strengths and motivations — growth paths per engineer, a high engineering bar, and consistently low attrition.

Scale & Realtime Systems

• Built the Classroom app from scratch across iOS and Android to 10M+ active users (16M+ total students reached) across the Middle East, Egypt, India and Pakistan.
• Built and led a 20+ engineer organization covering mobile, backend, realtime infrastructure and platform engineering.
• Realtime classroom system processing hundreds of millions of events per day — 50K+ concurrent WebRTC sessions at peak, ~6K events/sec sustained — across widely varying network conditions and geographies.
• WebRTC & Agora for realtime voice and interactive classrooms — solved synchronization between voice and sketch streams.
• CPU/GPU rearchitecture — all sketch rendering moved to the GPU (native iOS/Android graphics APIs), freeing the CPU for voice, chat, and network sync. CPU load during a live class: 90% → 20%. Battery drain per session: −40%. 2-hour classes ran cool with no heat throttling. State recovery after a network drop: under 2 seconds.
• Session replayability — independent voice and sketch event streams reconstructed and played back in sync on the user's device.
• Automatic sync framework using event fingerprinting and checksums — missing/delayed events identified per device and corrected without interrupting live sessions.
• Device-level observability — which user, teacher, session or geography was out of sync, not just aggregate vendor metrics; built on Apache Druid, Cassandra and event-processing infrastructure.
• Owned the reliability / latency / cost / experience trade-off at significant scale.

Leadership & Culture

• Individual growth and motivation first in a technically challenging environment — very low attrition across a 3+ year tenure.
• Worked closely with founders on product and technology direction; continued supporting the organization through the transition after leaving.

• End-to-end pharmacy and diagnostic test-booking platforms — product experience, backend services and integrations.
• Shipped customer-facing healthcare workflows end to end, expanding from mobile into broader backend and product engineering.

• Three-person founding engineering team — built a smart social-bookmarking product from the ground up across iOS and Java backend.
• Owned features end to end across mobile and backend — architecture, product and execution with limited resources.

• Java developer first, then native mobile across iOS and Android.
• Swift, Java, Kotlin, RxSwift, RxJava — consumer applications end to end.
• McDonald's (Europe/US) and Avon (UK) — shipped for global consumer brands.