MCS @ UIUC (GPA 4.0). Graduate researcher at SSAIL under Prof. Minjia Zhang. Before that, ~2.5 years at Oracle on telecom billing systems that handled 10M+ transactions a day.
I spend most of my time on LLM serving and RL post-training infra: speculative decode correctness, LoRA/weight-sync edge cases, and the places where train and serve disagree under load.
LinkedIn · Portfolio · bhatkar3@illinois.edu
SSAIL — GRPO post-training for agentic coding models on NCSA DeltaAI and AWS, with SWE-Bench Verified eval in Docker sandboxes. Recent debugging: an NCCL weight-sync hang that only showed up under certain topologies, and a fake 3.8× "speedup" that was really asymmetric GPU allocation.
Upstream contributions (vLLM, SGLang, verl) — correctness fixes on the serving/RL path:
- vLLM: reject chat turns the Jinja template silently drops instead of returning 200 with a truncated prompt
- vLLM: Model Runner V2 LoRA + speculative decode (logits / lm_head row alignment)
- vLLM: skip plain draft speculation past an independent draft model's
max_model_leninstead of CUDA-IMA'ing - verl: FSDP hybrid-engine sleep/wake sync for merged LoRA (level-1 path)
- SGLang: Falcon-H1 tied LM-head — avoid in-place
.float()on the shared head
Python · CUDA · PyTorch · vLLM · Ray Serve
Continuous batching + PagedAttention for Llama-3-8B. About 2.4× throughput vs naive batching, P99 under 800ms at 100+ concurrent requests. Multi-GPU Ray Serve with CUDA/KV profiling cut peak GPU memory ~22%.
Python · PyTorch · vLLM
Routes between single- and multi-token generation using entropy, margin, and n-best agreement. ~2.3× fewer decoding steps and ~28% lower long-form latency in the setups I measured. Batched LLM-as-judge cut the eval loop from ~90s to ~3.3s per sample.
C++ · Cassandra · PostgreSQL
Consistent hashing, 3× replication, recovery under simulated node failure. Indexed metadata + multithreaded ring lookups cut query latency ~40% in that setup.
TypeScript · React · Node · WebSockets · Redis
OT-based collab editor. Sub-100ms sync in testing, load-tested past 300 concurrent users. Tests focused on disconnect and concurrent-edit paths.
Python · PySpark · Kafka · MLflow
Streaming ETL + scoring at 10K+ txn/sec with sub-second micro-batch decisions.
Languages — Python, C++, Java, Go, Kotlin, TypeScript, Rust, SQL
ML / inference — PyTorch, vLLM, SGLang, TRL/GRPO, FSDP2, CUDA, Ray, TensorRT-LLM
Backend — Spring Boot, FastAPI, Node, gRPC, Kafka, Redis, Postgres, Cassandra
Infra — AWS, GCP, OCI, K8s, Docker, Terraform, Slurm/HPC, Prometheus, Grafana
Oracle — Software Engineer, Associate Consultant, Jul 2023 – Jan 2026 (Batelco Telecom, PPC Retail)
Eight Java/Kotlin microservices and 20+ REST APIs for subscription/billing/account flows at 10M+ daily transactions. Tracked down gRPC HTTP/2 stream exhaustion via Prometheus and retuned Istio pools (~37% higher throughput, ~12% lower P95 in that engagement). Led a 4-person team on a LangChain + Cohere docs assistant used by 500+ engineers.
Nibha Tech Solutions — AI Research Intern, Dec 2022 – May 2023
Feature engineering on 15GB+ / 25 years of market data; SARIMA + SHAP for forecasting under normal and stress regimes.
UIUC — MCS, GPA 4.0/4.0
VIT Pune — B.Tech Electronics & Telecommunication, CGPA 8.82/10
Oracle certs: Generative AI Professional, Data Science Professional, Machine Learning (Autonomous Database).
Chess (chess.com · lichess), powerlifting, cooking, drawing.
Email or open an issue if you want to talk about serving systems, RL infra, or anything above.