Skip to content
View alphasafal's full-sized avatar

Block or report alphasafal

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
alphasafal/README.md

Safal Gupta

AI systems engineer in the making. I build tools that expose how a system works: its state, failure modes, and evidence that it behaves as intended.

Computer Science at VIT Vellore · Based in India · Interested in AI infrastructure, evaluation, distributed systems, and robotics.

Animated terminal showing Safal's engineering workflow: build, evaluate, inspect, ship

Shipped

CONTRI — Open Source Compass · Chrome Web Store · Website

A Chrome extension that tells contributors, right on GitHub, whether an issue is really free to take (assignees, claims in the comments, linked pull requests), how quickly the maintainers reply, and whether the project is still active. Every number says what it covers, and missing data shows as unknown, never zero. No backend or tracking; optional read-only GitHub sign-in with the token encrypted at rest; 100+ unit tests plus Playwright end-to-end, accessibility and contrast checks in CI. Preact, TypeScript, Manifest V3, MIT licensed.

Open-source contribution

Merged (11 pull requests across Apple, Google, Microsoft, NVIDIA, AWS, Hugging Face, Sentry and others)

  • Apple coreai-optimization #143: quantizer config docstrings used \* escapes in normal strings, which raises SyntaxWarning on Python 3.12+ and stops the package compiling under python -W error. Made them raw strings, enabled ruff W605, and added a test that compiles every module with warnings treated as errors.
  • Google ADK #7726: the SKILL.md frontmatter parser split on the first --- anywhere in the file, which broke descriptions containing --- and dropped keys after block scalars. It now matches delimiters as whole lines. A maintainer reproduced the failures, approved it, and it landed on main through Google's Copybara import.
  • Microsoft Semantic Kernel #14661: ChatHistorySummarizationReducer with use_single_summary=False folded earlier summaries into each new one whenever the history began with a system prompt, the common case. Fixed where it finds the summary boundary and added a regression test. Approved by a maintainer and merged with all 34 checks passing.
  • NVIDIA NeMo Agent Toolkit #2467: the ReAct output parser cut a tool's input at any later word "observation", so tools received truncated JSON with no error and no retry. Fixed the parser and added regression tests, including multiline inputs.
  • AWS Powertools for Lambda (Python) #8669: EventBridgeV2Envelope raised UnicodeDecodeError on binary application/octet-stream payloads. Now tries JSON first and falls back to raw bytes. Bug confirmed by the maintainers; milestone 3.38.0.
  • Hugging Face Hub #5257: hf cache verify --local-dir reported the download metadata under .cache/huggingface/ as extra files, so --fail-on-extra-files failed even when every checksum matched. It now skips that metadata; added a regression test.
  • Sentry Python SDK #8119: serializing tracing baggage with no third-party items left a trailing comma, an empty list member under the W3C Baggage spec that strict downstream parsers can reject. Reachable through the Celery integration. A maintainer reproduced it with a Celery probe before merge.
  • CodexBar #4528: fixed cross-browser cookie permission diagnostics; added regression tests for error ordering.
  • OpenStatus #2908: fixed email subscriptions on self-hosted status pages; added regression tests. Merged after review.
  • Small fixes: corrected docs in Stanford Stanza #1697 and MIT CML blockly-plugins #91.

In review (each with a test or reproduction that fails before the fix)

  • Anthropic Python SDK #2021: the credential tests' environment isolation missed ANTHROPIC_BASE_URL, so with it set in the shell, 49 tests failed and 48 errored as token exchanges went to the wrong host. Approved by a reviewer who reproduced the before and after counts.
  • Swift Foundation #2364: the Gregorian calendar reported March 31, June 30 and September 30 as belonging to the next quarter. Now computes the quarter from the month. Approved by a maintainer.
  • AWS Powertools for Lambda (Python) #8670: the Kafka Protobuf deserializer read Confluent message indexes as plain varints instead of zigzag, silently dropping fields. Fixed the decoding and added parametrised tests across index shapes; milestone 3.40.0.
  • NVIDIA Warp #2159: array.transpose() accepted repeated axes and returned a view that read past the end of the array. It now rejects them, matching NumPy.
  • NVIDIA NeMo Curator #2534: the arXiv LaTeX cleaner deleted the character before every inline % comment and missed a trailing %. Fixed the pattern and added a regression test.

Open the source

Project What it explores
AI Eval Engineering A TypeScript evaluation harness for prompts and agents, with graders, trajectory metrics, cost and latency tracking, and release gates.
Production Agent Skills Eight portable engineering, data, and product skills with validation and evaluation cases.
Exchange Core A C++20 limit order book with pre-trade risk, a write-ahead journal, deterministic replay, and network gateways.
SAAYA VERIFY An evidence-led cooling shelter research concept and tabletop fabrication candidate.

Current focus

  • AI systems: evaluation, retrieval, agents, and efficient inference.
  • Systems engineering: deterministic behavior, observability, storage, and distributed systems.
  • Robotics: connecting software decisions to the physical world.

I learn by building the underlying pieces, testing them, and making the tradeoffs visible. My recent work spans Python, C++, TypeScript, and full-stack tools.

Connect

Portfolio · LinkedIn · X · HackerRank

Pinned Loading

  1. contri contri Public

    🧭 Find open-source issues you can actually take — see if a repo is alive, if maintainers reply, and if an issue is already claimed. Right on GitHub.

    TypeScript 4 2

  2. ai-eval-engineering ai-eval-engineering Public

    A practical TypeScript AI evaluation harness for testing prompts and agents with deterministic graders, trajectory metrics, cost/latency tracking and release gates.

    TypeScript

  3. openstatusHQ/openstatus openstatusHQ/openstatus Public

    Open-source status page, incident response and uptime monitoring

    TypeScript 9.3k 809

  4. saaya-verify saaya-verify Public

    SAAYA VERIFY: Vellore cooling-shelter research lab, interactive evidence demo and tabletop fabrication candidate

    Python

  5. steipete/CodexBar steipete/CodexBar Public

    Show usage stats for OpenAI Codex and Claude Code, without having to login.

    Swift 23.1k 2.2k