Sarvex Jatasra

Lead engineering at a lab that builds the reinforcement-learning environments and human-in-the-loop data infrastructure that frontier model teams use to post-train and align their models. Championed deterministic, reward-hacking-resistant evaluation rewards tied to published benchmark metrics with no model in the reward path to keep training signals trustworthy at scale. Own the technical direction for verifiable RL environments and data pipelines spanning code, audio, video, reasoning, and security domains. Owned engineering hiring and the technical bar as the company scaled to ~1,200 people across engineering, operations, and data teams, building the org structure and interview pipeline to sustain that growth. Established the evaluation and data infrastructure underpinning the labs frontier-model alignment work for 4 frontier-model teams dozens of environment families generating 20,000+ verifiable task instances, with a pipeline producing ~20M verified samples per week taking it from research prototype to production-grade infrastructure.

Owned end-to-end technology for a fintech platform built on LLMs and Retrieval-Augmented Generation. Led architecture, engineering hiring, and delivery for a ~20-person engineering organization. Built the data pipelines that powered customer segmentation and sharpened product targeting, measurably improving conversion on targeted segments versus the prior broad-targeting baseline and turning raw data into actionable go-to-market signals. Reduced infrastructure cost and improved delivery velocity by consolidating onto a managed multi-cloud (AWS / GCP / Azure) MLOps setup for reliable model deployment and monitoring.

Led solution architecture across multiple engineering pods at a large technology organization, owning both cost discipline and applied-AI product delivery. Cut monthly AWS spend from ~$500,000 to ~$90,000 (82% reduction) through a focused FinOps program right-sizing, workload consolidation, and reserved/committed-use planning while holding p99 latency and availability flat, with no customer-facing regression. Took BUDDY from prototype to a deployed enterprise assistant—a transformer-based voice assistant with speech recognition and generation and vernacular Indian-language support, integrated with OpenAI / GPT-3 through a multi-tier processing architecture. Set architecture standards and reviewed designs across pods, balancing delivery speed against long-term maintainability and cost.

Co-founded a clean-energy startup and built the engineering organization from the ground up. Cut solar-customer onboarding from roughly six months to a couple of hours across thousands of auto-generated proposals per month by replacing manual rooftop site surveys with a machine-learning pipeline over satellite imagery that computed usable rooftop area within 5% of manual site surveys (a Project Sunroof-style approach). Built the web and mobile platforms end to end, owning architecture, delivery, and the engineering team. Integrated Salesforce CRM to drive sales and Power Purchase Agreement (PPA) workflows, connecting the technical pipeline to the business process.

Led mobile engineering across platforms during a period of rapid product growth. Grew the mobile app from under 100 to 10,000+ daily active users in ~9 months at 99% crash-free sessions, earning Play Store Editors Choice and App Store featuring. Cut feature delivery cycle time from ~six months to ~two weeks by restructuring teams and release process. Stood up dedicated Android, iOS, and Windows teams across ~50 engineers, and introduced THOR, a middleware layer that improved app responsiveness.

Built Amazon Prime Video client applications from inception across multiple device platforms. Delivered the V1 client platform at Amazon scale with adaptive-bitrate streaming tuned for networks from 2G through 4G/LTE, managing a two-pizza engineering team (~12) across hiring, delivery, and operational quality.