Prefer plain text? Scroll down ↓
Hi, I'm Kesh
 
 
 
Ukeshwaran Muthukumaran Radha (Kesh Muthu)

Open to internships and 2027 new-grad roles

Kesh
Now
Informatics at UW

Open to internships and 2027 new-grad roles. Software, AI, security and networking.

Now
 
 
AAplyxJob search now takes 0.5s, down from 7snow
RRipprzFixed 3 Stripe payout race conditions. No more double refunds.2m
Swipe up to open

Ukeshwaran Muthukumaran Radha (Kesh Muthu)

Software, AI, security and networking. Informatics at the University of Washington. Open to internships and 2027 new-grad roles.

keshm@uw.eduGitHubLinkedIn

Experience

KredosAI

Voice agent, 2026 to now
  • FastAPI and vLLM microservices serving quantized LLMs on GPU, 100+ calls a day
  • Tuned KV cache and request batching to stay under 500ms
  • Async Python handles 10 simultaneous calls without added lag; input sanitizing cut prompt-injection vectors 45%
  • Moved from LiveKit Cloud to a self-hosted stack so customer audio stays in-house
CallerVoice agentvLLM on GPUReply

KredosAI

SMS support agent, 2025
  • Fine-tuned LLMs with LoRA/QLoRA over 30+ runs, response quality up 26%
  • Filtered RAG output to block PII leaks, incidents down 73%
  • Trained custom embeddings in ChromaDB: relevance up 30%, speed up 20%
  • Added RCS next to SMS through Twilio, with failed RCS webhooks rerouted to SMS

Apexiel

AI/ML intern, 2024
  • Semantic search API with SentenceTransformers: precision up 25%, lookups 35% faster
  • Parallel pipeline chunking and embedding PDFs into Qdrant, ingestion 2x faster
  • Scaled to thousands of documents with lookups inside 4 seconds

Projects

Ripprz

10s to 13ms
  • Pokemon card trading app. Led 5 developers: standups 3x a week, code review, and I owned payments
  • Fixed 3 Stripe payout race conditions with lock ordering and idempotency keys, ending double refunds
  • Inventory allocation with prefix sums: 10s down to 13ms
  • Return-to-back draw queue cut repeat draws from 44% to 0% in a 240-pull test
AppAPILock + idempotencyStripe

KTP platform

50+ members
  • Chapter site that replaced manual dues tracking, event turnout up 50%
  • Stripe webhooks across 4 payment types, errors down 90%
  • Postgres row-level security stops members promoting themselves to admin
  • Retries failed dues emails on a cron, on-time payments up 20%

Aplyx

7s to 0.5s
  • Job application agent: searches Ashby, Lever, Greenhouse and Workday, tailors a resume per posting, applies
  • Read-only MCP server so other AI agents can query job data
  • Redis caching took searches from 7s to under 500ms
  • Google OAuth, bcrypt, Supabase RLS, and 2FA on new devices
SearchDedupeTailorApplyDiscord

School networks

8 schools
  • Led a team of 5 deploying Cisco routers and switches with OSPF and DHCP
  • VLANs for device separation, plus subnet and IP plans for each site
  • Found DHCP scope exhaustion with Wireshark and fixed a site-wide outage
  • Supported the rollout of 1,200+ laptops across the district
router: OSPF + DHCPswitch: VLAN 10switch: VLAN 20switch: VLAN 30

Password manager

security
  • HTTPS with TLS and parameterized queries: no plaintext exposure, no SQL injection
  • Port scans and enumeration to find exposed services
  • OWASP ZAP scans to confirm each fix held
  • Also used as a teaching codebase with planted bugs for UW's INFO 310

Home lab

50+ users
  • Self-hosted services on Ubuntu with automated deploys and monitoring
  • Port forwarding, static routing and DNS on an isolated network
  • SSH key tunnels for remote access, services split across VirtualBox VMs

Skills

Languages
Python, TypeScript, JavaScript, Java, SQL, C++, Rust, Go (learning)
AI
vLLM, LoRA/QLoRA, RAG, MCP, Qdrant, ChromaDB
Web and cloud
FastAPI, Next.js, React, PostgreSQL, Redis, Supabase, Stripe, AWS, Docker, Linux
Security
OWASP, RLS, OAuth, Security+ in progress
Networking
CCNA, TCP/IP, OSPF, VLANs, Wireshark

Contact

Email keshm@uw.edu, or find me on LinkedIn and GitHub.