$ whoami
ishaq ahmed shaik
mle / swe · san francisco, ca
ms cs @ asu, graduated may 2026 · gpa 3.8
$ looking_for
full-time roles
primary mle / ml infra / ai research eng
also open backend / infra / distributed systems swe
us (remote or on-site) · us work authorized
$ links
resume [ml.pdf] [backend.pdf]
email ishaq.sde@gmail.com
github sonuishaq67
linkedin si67
$ cat recent.md
[2026-04] won villagehacks 2026 @ asu
with benjamin aspinall. built agent debate: a multi-agent research orchestrator chaining 5 agents (plan, retrieve, summarize, critic, write) into grounded opposing reports, then a podcast-style debate via elevenlabs tts, tracking per-turn claim confidence with the beliefs sdk.
$ experience
[25-26] asu, graduate student assistant
deployed a rag chatbot to 5k+ users at 1.2s latency on aws lambda; cut multi-turn compute costs 40% with context summarization. scaled a lambda + rds ml pipeline to 300k+ events at 99.9% uptime.
[23-24] extreme networks, software engineer
migrated java rest apis to grpc + protobuf: 25% inter-service latency drop across 1m+ daily requests. shipped 8 spring boot data apis across 20+ nfl/mlb stadium sites for flagship hardware.
[2023] nbyula, swe intern
cut image load 84% (2.5s → 400ms) with cloudfront edge caching + server-side cropping; raised chat dau 32% by fixing redis cache invalidation.
[2022] loadshare networks, sde intern
dockerized spring boot microservices with github ci/cd — 75% faster deploys (40 → 10 min). built a slack bot that automated manual sql ops.
$ things_i_built_for_fun
chess transformer on lichess
500m-param decoder-only transformer trained on 10b tokens across 8xgaudi2, fed by a duckdb pipeline over 100m+ games. serves live games at sub-second latency via onnx on ec2. challenge it on lichess.
→ github.com/sonuishaq67/chess
pong, but the agent learns it
double-dqn reinforcement learner that hits 100% win rate on 10k games. 91% training speedup from 16 parallel async vector envs + torch.compile.
amazon reviews rag
hybrid vector + knowledge-graph retrieval (pgvector + neo4j) over 571m amazon reviews. +27% precision@10, with ablations against vector-only and graph-only baselines.
network intrusion detection
xgboost classifier on nsl-kdd flagging malicious connections; cut the input space 56% via correlation filtering + gain importance with no accuracy loss.