Skip to content
View TanishGudise's full-sized avatar
0 followers · 1 following

Achievements

Achievement: Pull SharkAchievement: Pair Extraordinaire

Achievements

Achievement: Pull SharkAchievement: Pair Extraordinaire

Block or report TanishGudise

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. 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
TanishGudise/README.md

Tanish Gudise

CS + Business @ Carnegie Mellon '27 · GPU kernels and inference, agent infrastructure, and evals

  • Megakernels #15: found a stale read in Hazy Research's Llama-1B megakernel on H100. 26 of 1,000 stock generations produced a wrong token, vs 0 of 1,000 with the fix.
  • ThunderKittens #215: bind_kernel returned silently when a launch failed, leaving outputs unwritten; it now raises the CUDA error.
  • daft-ai-audit · Daft #7643: 11 bugs in Daft's AI functions under fault injection, including silent row loss with vLLM on Ray; fixes for the top three.
  • unsloth-zoo #2095: MoE QLoRA replay pin on by default, 1.063x faster.
  • silent-failure-bench · session-memory-bench · claims-vs-reality: pre-registered, exact-graded benchmarks for agent silent failures, cross-session memory, and agents' self-reports.
  • pipeline-reliability: 282 validated Harbor tasks in 56 problem families where agents fix reliability bugs in LLM data-pipeline code, each passing an automated QA gate.
  • fin-model-edits · filing-update-evals: exact-graded agent evals on real SEC filings and credit agreements; no LLM judge.
  • Merged reliability fixes in agent infrastructure: Composio (co-author), Cua, Daytona, Osmosis.
  • Airbnb, AI Infrastructure intern: built a conversational agent for an internal prompt-engineering platform, with RAG, OpenTelemetry tracing and an eval harness. Full-time return offer.
  • Parameter Golf: 16MB language model trained on 8×H100s with 60+ tracked ablations; briefly topped the leaderboard.

Python · C/C++ · CUDA · Triton · Go · TypeScript · PyTorch · vLLM · Ray · Docker

Open to full-time roles starting January 2027 · Website · LinkedIn · tgudise@andrew.cmu.edu

Pinned Loading

  1. daft-ai-audit daft-ai-audit Public

    Fault-injection harness for Daft's AI functions: one command reproduces 11 bugs (fake vLLM/OpenAI backends, local Ray, no GPU or API keys)

    Python

  2. silent-failure-bench silent-failure-bench Public

    Exactly labeled silent failures in a tool-using agent: detectors and a fix-trust gate

    Python

  3. claims-vs-reality claims-vs-reality Public

    Pre-registered eval: do AI agents misreport what they did? Exact ledger-based detectors vs LLM transcript judges.

    Python

  4. session-memory-bench session-memory-bench Public

    Which memory from earlier agent sessions helps the next coding session, and when it hurts: a pre-registered benchmark graded by hidden tests

    Python

  5. pipeline-reliability pipeline-reliability Public

    A generator of 282 validated Harbor tasks for fixing reliability bugs in LLM data-pipeline code, with an automated QA gate

    Python

  6. parameter-golf-16mb parameter-golf-16mb Public

    16MB LM for OpenAI Parameter Golf — 1.05670 BPB, 8×H100

    Python