Skip to content
View devtyagi3909's full-sized avatar
:shipit:
bankai
:shipit:
bankai
11 followers · 12 following

Achievements

Achievement: QuickdrawAchievement: Pull Sharkx2Achievement: YOLO

Achievements

Achievement: QuickdrawAchievement: Pull Sharkx2Achievement: YOLO

Block or report devtyagi3909

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.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. 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
devtyagi3909/README.md

dev tyagi // applied ai engineer

applied ai telemetry

divider

upstream merged

unsloth zoo pr 2290 vllm pr 56635 truss pr 2682 cva6 pr 3518 riscv-dbg pr 206 riscv-dbg pr 204

  • unslothai/unsloth-zoo #2290: Precomputed Transformers v5 check in GptOssTopKRouter forward loop to eliminate runtime version parsing overhead in MoE routing.
  • vllm-project/vllm #56635: Isolated reasoning boundary tokens (<think>, </think>) in streaming ParserEngine, eliminating tool-call argument contamination on reasoning models.
  • basetenlabs/truss #2682: Upstream runtime model serialization and container packaging fix.
  • openhwfoundation/cva6 #3518: Instruction pipeline verification and testbench retirement sequence fixes.

// flagship projects

support-triage-mcp tool-call-bench factforge

project focus & problem verified evals & metrics live links
support-triage-mcp forward-deployed style triage agent for messy customer support data (emails, webhooks) with MCP server tools (CRM lookup, Postgres query, Stripe refund gate) 99.4% schema adherence · 91.4% intent routing precision · sub-2s latency · Langfuse audit trail pipeline server · eval harness · test corpus
tool-call-bench benchmark evaluating tool-calling & structured-output reliability across providers (schema violations, malformed arguments, boundary leaks, retries) 0.8% failure on sonnet-3.5 vs 7.4% on gpt-4o-mini · measured token retry overhead · benchmark harness via LiteLLM results table · repro harness · blog post
decodeforge high-throughput grammar compiler & constrained decoding runtime for production LLM serving pipelines O(1) sub-token bitmask filtering · sub-microsecond logit pruning · zero GPU decode stalls architecture · spec · quickstart
factforge automated fact extraction, source PDF grounding & cross-document contradiction detection for financial filings 100.0% verbatim excerpt grounding · 94.2% token reduction vs pairwise prompting · O(M log N) incremental retrieval architecture · eval report · api server

// in review (upstream)

sglang pr 47534 sglang pr 47506 vllm pr 64716 sglang pr 47470 pytorch ao pr 5041 vllm pr 64730 sglang pr 45031 sglang pr 44571 transformers pr 49614

// case study

debugging model tool-call parsing in vllm (vllm-project/vllm #56635): models generating reasoning traces before tool calls often emit boundary tokens (<think>, </think>) that leak into raw parser buffers, causing structured tool extractors to fail or emit invalid JSON payloads. identified the root cause in GLM frontend token boundaries, upstreamed the fix directly into the base ParserEngine, and added test assertions verifying zero boundary token leakage during streaming tool-call extraction.

// stack

• languages  : python, typescript, c++, sql, bash
• genai/llm  : litellm, langgraph, openai/anthropic sdk, mcp (model context protocol), langfuse
• data & db  : postgresql, pgvector, pymupdf, rapidfuzz, networkx
• backend    : fastapi, uvicorn, pydantic, docker, kubernetes, aws (ec2, s3)
• evals/qa   : pytest, promptfoo, ruff, git, postman

• also: hardware background — synthesizable 4-wide out-of-order rv32i core (1st place @ sandisk hackathon), fpga zynq-7000 dma, and uvm verification.

portfolio github email linkedin

Pinned Loading

  1. vllm-project/vllm vllm-project/vllm Public

    A high-throughput and memory-efficient inference and serving engine for LLMs

    Python 96.4k 24.5k

  2. openhwfoundation/cva6 openhwfoundation/cva6 Public

    The CORE-V CVA6 is a highly configurable, 6-stage RISC-V core for both application and embedded applications. Application class configurations are capable of booting Linux.

    Assembly 3.2k 1k

  3. riscv-ooo-core riscv-ooo-core Public

    Full superscalar RV32I out-of-order processor with Tomasulos Algorithm, 7-stage pipeline, and UVM testbench. 1st Place + Special Jury Award — SanDisk Hardware Hackathon.

    Verilog 13

  4. constraintforge constraintforge Public

    Production-grade FPGA timing constraint library for Xilinx UltraScale+. Modular, parametric XDC templates for high-speed digital design.

    Tcl 4 1

  5. huggingface/transformers huggingface/transformers Public

    🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.

    Python 169k 35.1k

  6. sgl-project/sglang sgl-project/sglang Public

    SGLang is a high-performance serving framework for large language models and multimodal models.

    Python 38.5k 10.2k