experience

software engineering intern · menlo park, ca

• ·selected as 1 of 10 fellows from Central Asia; worked alongside the C-suite to ship "Company Brain" AI assistants to small businesses
• ·rebuilt the LibreChat UI (React 18, Vite, Tailwind, Radix) into a white-label product live for 6 SMB clients: a d3-force knowledge-graph explorer, a Recharts analytics dashboard, and an approval queue over new Express + MongoDB APIs
• ·cut LLM costs 33% by routing simpler tasks to cheaper models in a two-stage ingestion pipeline (triage, then distillation via OpenRouter) that distills chats into a wikilinked Markdown vault backed by pgvector
• ·provisioned 15+ Linux VPSs with Docker; spent 40+ hours onboarding clients onto white-labeled deployments

systems administrator · vice president · swarthmore, pa

• ·run production infrastructure for a 1,600-student campus at 99% uptime: a 128-thread, 503 GiB Proxmox hypervisor with 15 VMs and a 94-container Docker Swarm behind Traefik, with Keycloak SSO in front of everything
• ·serve an RTX 6000 Pro (96 GB) and an RTX 5090 (32 GB) to 40+ professors and students for research, vLLM inference, and 3D rendering; 70 TB of ZFS with Proxmox Backup Server, Prometheus/Grafana, and CrowdSec
• ·as vice president since feb 2026, lead the largest and best-funded student tech org on campus (100+ members): member project teams, a $15,000/semester budget, recruitment, events, and operations

founder & lead · swarthmore, pa

• ·founding Swarthmore's first AI safety student group, funded by Kairos
• ·launching an 8-week AI Safety Fundamentals reading group and a speaker and workshop series on alignment, interpretability, and AI policy for fall 2026

mentee · remote

• ·1 of 250 selected from 1,500 applicants for a 10-week 1:1 mentorship with Google SWE Parneet Kaur: C++, data structures and algorithms, and technical interview prep

undergraduate researcher · swarthmore, pa

• ·recovered up to 36 unknown ODE parameters of a colorectal-cancer signaling model (2× the autodiff baseline) by training Fourier-feature PINNs (Adam, then L-BFGS) with a derivative-free trapezoidal integral residual
• ·triangulated parameter identifiability via Fisher information, profile likelihoods, and Sobol sensitivity (SALib); ported a 27-species ODE model from MATLAB to Python at <1% error on a 64-core queued GPU platform
• ·paper accepted to the SIAM conference in NYC; awarded a competitive $6,000 internal research grant to fund the project

lead swe & pm · swarthmore, pa

• ·adapted the open-source LibreChat stack into Swarthmore's fully self-hosted campus LLM, serving 1,300 unique users (200 weekly active, 88M tokens/week) with Qwen3.6-35B-A3B (MoE, FP8) via vLLM; no data leaves campus
• ·replaced LibreChat's stock RAG with a custom TypeScript hybrid retrieval engine (parallel Qdrant dense + sparse lexical search, TEI cross-encoder reranking, hard 1.5s fail-open budget); 96% answer accuracy, average latency from 4.5s to 1.8s
• ·wrote the Python ingestion pipeline: 3,528 campus docs into 11,302 chunks in 23s, with an atomic alias swap for zero-downtime reindexing
• ·own the product roadmap and feature prioritization as PM

tech fellow (sp26 cohort) · remote

• ·selected to the tech track of a competitive fellowship (sub-8% acceptance) focused on engineering skills and technical career prep
• ·sharpened engineering and professional skills through mentorship and direct engagement with corporate partners

algorithmic research · remote

• ·researched the NP-hard Graph Bandwidth Problem on a 100% scholarship with prof. David Perkins
• ·beat GRASP, simulated annealing, genetic algorithms, and ICA by ~20% with a novel genetic algorithm using wave function collapse (NGA-WFC)
• ·published as an Outstanding Research Paper in the Pioneer Journal (top 1.3% of submissions)

founder & ceo · kazakhstan

• ·secured $30,000+ from Samsung Innovations, UNESCO Startups, and Kazakhstan's Ministry of Ecology to tackle large-scale bee population decline
• ·grew to 300+ active beekeepers with a data-driven dashboard in React, TypeScript, Next.js, and MongoDB
• ·delivered real-time hive-health predictions with an ML pipeline over 500,000+ datapoints — Confluent Kafka, LightGBM, MQTT, TimescaleDB, FastAPI
• ·reached $1,000 MRR with a live apiary-recommendation system over LoRaWAN powered by a fine-tuned Mistral-7B-Instruct-v0.3