Skip to content
View buddywhitman's full-sized avatar

Organizations

@ManipalHackathon2024

Block or report buddywhitman

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
buddywhitman/README.md

Systems engineer working at the intersection of AI inference infrastructure and safety-critical embedded systems.

Currently a pre-final year EE undergrad at MIT Manipal - published in Springer Nature Q1 (2026), IEEE VTC (2026), and 3 provisional patents through MAHE/DGIP.

Building multilingual voice AI @ Fettle · solar racing EV firmware @ SolarMobil · agentic B2B SaaS @ Guaq AI


Open Source Contributions

All PRs below are currently in review - code-owner responses pending.

Repo PR What it does
vllm-project/vllm #44794 FlowPrefill: checkpoint-based sub-chunk preemption for the v1 scheduler — CUDA event polling detects decode pressure and suspends prefill mid-chunk without synchronize(), using a pre-allocated event pool to eliminate per-step overhead
sgl-project/sglang #27498 P-EAGLE: parallel speculative decoding for SGLang — generates all K draft tokens in a single forward pass instead of K sequential passes, porting the technique from vLLM v0.16.0 and closing #23171. +37K lines, 14 review comments
microsoft/vidur #80 MoE + disaggregated scheduler: DeepSeek-V3 and Mixtral-8x7B MoE configs with BaseMoEModelConfig, plus a disaggregated prefill/decode scheduler for Microsoft Research's LLM inference simulator
anthropics/anthropic-sdk-python #1663 Streaming perf: skip build_events() and use O(1) buffers on the get_final_message drain path — eliminates unnecessary TextEvent/InputJsonEvent construction when caller doesn't iterate events, fixing #1649
jpmorganchase/QOKit #139 Lazy Ising kernel: on-the-fly Ising energy computation inside the CUDA phase-separation kernel, eliminating the O(2ⁿ) precomputed diagonal array and its CPU→GPU transfer, fixing #35
finos/pylegend #311 Type system: adds Time, Variant, and Binary primitive types through the full type chain (enum → class hierarchy → expression layer) in FINOS/JPMorgan's Python Legend query library, closing #220

Research & IP

📄 Springer Nature Q1Technical Validation of a Multimodal Emotion-Adaptive Biofeedback System for Autonomic Regulation · Scientific Reports, Apr 2026
📄 IEEE VTC 2026Physics-Informed Stochastic Receding Horizon Control for Autonomous Energy Management in Solar Racing · Accepted Apr 2026
🔒 3 provisional patents via MAHE/DGIP — adaptive EV interface (202541050036), long-range EV telemetry (202541064383), robotic solar panel cleaning (202541079610)

Stack

AI/Inference    vLLM · SGLang · PyTorch · TorchRec · LangGraph · LangChain · Gemini ADK
Voice/RT        LiveKit · SIP/PSTN · WebSockets · Whisper · multilingual STT/TTS
Backend         FastAPI · Django · Spring Boot · Redis · Kafka · PostgreSQL · ClickHouse
Platform        Lago · Clerk · PostHog · Cloudflare Zero Trust · Prometheus · Grafana
Embedded        STM32H7 · FreeRTOS · TouchGFX · CAN · MISRA C/C++ · Verilog · TensorRT
Cloud           AWS (ECS, Fargate, Lambda) · GCP (Cloud Run, Vertex AI) · Kubernetes

Current Areas

  • AI inference serving — speculative decoding, scheduler preemption, MoE simulation
  • Voice AI infrastructure — multilingual hospital workflow agents, SIP orchestration, inference pooling
  • Safety-critical firmware — Cortex-M7 ECU, FreeRTOS, CAN bus, race strategy digital twin
  • Agentic platform primitives — multi-tenant auth, metering, tool integration, observability

Contact

📧 pulkit.talks@gmail.com · university
🌐 buddywhitman.tech · LinkedIn · ORCiD
💬 Discord: buddywhitman#1784

Pinned Loading

  1. sglang sglang Public

    Forked from sgl-project/sglang

    P-EAGLE parallel speculative decoding for SGLang - upstream PR #27498

    Python

  2. vllm vllm Public

    Forked from vllm-project/vllm

    implementing FlowPrefill - checkpoint-based sub-chunk preemption for v1 scheduler (upstream PR #44794)

    Python

  3. solarmobil/stm32h750b-dk solarmobil/stm32h750b-dk Public

    Primary ECU & Driver Interface Firmware

    C

  4. dist-gcs-pdf-processing dist-gcs-pdf-processing Public

    Distributed Multilingual Document Processing System Built for XX-TB Scale

    Python 1 1

  5. strategy strategy Public

    Real-Time Receding Horizon Model Predictive Control (MPC) & Physics-Informed (PINN) Digital Twin

    Python

  6. aqi-informatics-india aqi-informatics-india Public

    Environmental Intelligence System for the Indian Subcontinent

    Python