𓆝 𓆟 𓆞anush sonone𓆞 𓆟 𓆝

@pwc - fall 2026, incoming ai engineering intern on the tide (tax innovation & delivery experience) team.

@visa - cut onboarding from 3 mos to 3 days by building langgraph workflows to auto-provision users from admin reqs; achieved 90%+ service grant accuracy by architecting a postgres-backed knowledge graph for entitlement routing; and drove notif loss to near-zero across 16k+ provisioning events by implementing redis streams with retries & dlq.

@afterquery (yc w25) - accelerated retraining cycles by 20% across 10 domains by designing agent eval pipelines for manus ai tool-calling; lifted tool-call accuracy by 18% at a 2m+ user scale using python harnesses on deterministic truths & diffs; and replayed 10k+ bad tool calls without re-hitting prod by shipping append-only failure-mode stores in postgres/s3.

@profluento ai - raised $100k; engineered a lead scoring engine to surface high-net-worth prospects for wealth management firms; onboarded 15+ firms by scoring linkedin users with a node.js weighted-matrix model over 10+ normalized signals; and reduced paid profile enrichments by 80% using async job routing on cloudflare queues with regional kv caching.

@cdk global - enabled full iam control across 44k+ dealerships by leading implementation of cdk iam's "super admin" role; reduced unauthorized access to near-zero in pilot through per-request auth built with node.js & dynamodb; and eliminated admin-audit gaps across 10k+ updates/week by integrating iam audit logging & change tracking.

killmycluster: a self-healing raft-based kv - wrote a raft key-value store from scratch in go with a crash-safe wal & grpc replication, self-healing from leader loss in ~10s; cut read cost by moving gets off the raft log onto readindex leases; and sustained ~1.2k replicated writes/s & ~1.5k reads/s on a 4-ocpu box while visitors kill machines on purpose at anush.wiki/blog/raft.

lapsynk: formula 1 ai copilot - placed 3rd at hacktx 2025 (733 hackers) with a telemetry-driven message prioritization engine for race engineers; and delayed non-critical msgs up to 5s by implementing message deferral across traffic density, g-force load, & braking.