Skip to content
View Danishlynx's full-sized avatar

Block or report Danishlynx

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
Danishlynx/README.md

Danish Ali

Software Engineer · Manchester, UK
3rd place, AWS Trainium Frontier Competition (1st in Round 1) · 1st in the qualifiers, AMD × GPU MODE E2E Model Speedrun
DataHub contributor · MSc Advanced CS, University of Manchester

LinkedIn Email


I work on ML training and inference performance, and I build systems that check their own work: deterministic simulations with server-side replay, allocation stress-tested for zero oversell, agent output gated by linters and real browsers. Previously at Energy Web and AAK Tele-Science. Open to software engineering and ML performance roles in the UK.

ML performance

Name What I built Stack Links
AWS Trainium Frontier Competition Training a language model from random weights on Trainium2 chips inside a fixed time budget. 3rd overall of 100 teams, and 1st in Round 1. Round 1 (1 chip, 30 minutes): 0.9328 bits per byte, with throughput up from about 42k to 340k tokens a second, mostly from custom NKI kernels (attention, the loss head, fused elementwise blocks) and fp8 matmuls. Round 2 (16 chips, 4 hours): a 3.68B-parameter mixture-of-experts model (0.71B active per token), step time down from the kit's 2.41 s to 0.86 s, and the lowest validation loss of the ten finalists. NKI · Neuron SDK · PyTorch · Python Repo
AMD × GPU MODE E2E Model Speedrun Qualifiers: 1st overall, with a fused MoE kernel at 69.9 µs geomean on the MXFP4 track. Finals: DeepSeek-R1 serving on 4x MI355X, time per output token down from 6.30 ms to 5.64 ms, 2 of 4 gates passed. Triton · ROCm/HIP · AITER · FlyDSL · Python Kernels · Finals

Open source: DataHub core

Contributor to DataHub (12k+ stars), the open-source metadata platform. Designed and shipped incident support for ML entities and keyless Snowflake authentication, with 8 PRs merged and 2 approved across the metadata model, GraphQL API, React UI and Python ingestion connectors.

Status Contribution
Accepted RFC #18911: per-layer analysis of how entity support drifts, adopted as the maintainers' "gates" roadmap model
Merged #19112 · #19132 · #19367: incident support for ML models, features and feature tables across the metadata model, GraphQL API, authorization policies, React UI and MCP tooling
Merged #19889: keyless Snowflake login through workload identity federation, so ingestion on AWS, Azure or GCP uses its own cloud identity and stores no password or key. Verified live on AWS
Merged #19539: a single undecryptable secret no longer fails the whole batch, backported by maintainers into the v1.6.0.3 release
Merged #19540: structured property values containing a URN no longer crash the entity page
Merged #19278 · #19537: a CI flake fix and API docs corrections
Approved #19627: health status for ML entities across lineage, search and entity pages
Approved #19879: platform instance support for Superset and Preset dashboards, charts and datasets
In review #19629: a CI test that fails when incident support disagrees between the write, summary and GraphQL read layers, picked up at a maintainer's request
In review #19535 · #19536: deterministic replacements for two timing-based CI tests, root-caused from CI logs

Maintainer, on closing the RFC: "We now have full support for incidents on these 3 entities across data model, API, UI, and MCP."

Snowflake auth maintainer, approving #19889: "I broke each new guard on purpose again [...], and a test caught every one."

Maintainer, approving #19540: "nice fix and a really clear writeup."

Projects

Name What I built Stack Links
CivicNexus Governed agent fleet for municipal permit casework. Four agents on Vertex AI read the application, retrieve the municipal code, and draft a determination whose every quote is byte-checked against the committed statute. Email to human gate in about 62 seconds. Attachments are OCR'd and re-screened before any model sees them, and only a named human can issue a permit, enforced by a write-once approval row. Vertex AI Agent Engine · ADK · Gemini · Gemma · Cloud Run · Terraform Live · Repo
Blast Radius ML supply-chain agent on the DataHub metadata graph. Finds every model in the blast radius of an upstream schema change, files evidence-backed incidents, opens dbt-validated fix PRs, and blocks bad deploys through a CI circuit breaker. Building this surfaced the gap behind the RFC above. Python · DataHub · MCP · dbt Repo
Singleton No-oversell allocation of scarce slots. Sharded counters under optimistic concurrency and a commit-reveal lottery anyone can re-verify in the browser. 10,000 concurrent claims on a 200-slot release, zero oversold. Aurora DSQL · Next.js 15 · TypeScript · Vercel Live
GPTDesign Design-system-locked UI generation for coding agents. A deterministic token-drift linter hard-rejects off-system output, a sight loop has the model inspect its own render, and a Playwright gate clicks every tab, dialog and button before a run passes. 880+ tests. TypeScript · Fastify · Playwright · MCP · SSE Repo
Precedent Decision-memory agent for Slack. Extracts decisions from threads with confidence gating, keeps a forward-only supersession history, and answers "what did we decide about X?" with source-thread receipts. Exposed as an MCP server. Slack Bolt · Claude · MCP · SQLite · Railway Live · Repo
underVault Asymmetric-knowledge roguelike on Reddit. One shared dungeon, one life a day, rules hidden from you and learnable only from other players' corpses. Integer-only deterministic simulation with server-side replay validation. Devvit · Phaser · Hono · Redis Repo
Earlier projects
Name What I built Stack Links
Web Information Decay Quantifying the degradation of web archival quality across modern platform architectures Python · Playwright · Wayback Machine Repo
Ethereum Charts Real-time chain analytics dashboard React · Web3.js · D3 Live
Proxy Wallet Gas-optimised smart wallet with guardian recovery Solidity · Hardhat · ethers.js Live

Stack

Languages TypeScript · JavaScript · Python · SQL · Solidity
ML performance NKI (AWS Trainium) · Triton · ROCm/HIP · FP8 and MXFP4 · PyTorch
Backend Node.js · Fastify · Express · FastAPI · GraphQL · REST · WebSockets
Data PostgreSQL · Aurora DSQL · Redis · MongoDB
AI & agents LLM APIs (OpenAI, Anthropic) · MCP servers · agent evals · Playwright automation
Frontend React · Next.js · Tailwind CSS
Cloud & DevOps AWS · Docker · Kubernetes · Terraform · GitHub Actions

Pinned Loading

  1. blast-radius blast-radius Public

    Autonomous ML supply chain guardian on the DataHub metadata graph, finds every model in the blast radius of an upstream change, files evidence-backed incidents, blocks bad deploys, and remembers th…

    Python 1

  2. GPT_Design GPT_Design Public

    The design layer for Codex. You type what you want. An AI agent designs it live on a canvas, in front of you. But unlike normal AI generation, the design is locked to a design system: the exact col…

    TypeScript

  3. precedent precedent Public

    Decision memory for Slack. Watches team channels, extracts decisions with Claude, stores them with permalinks, answers "what did we decide about X?" in the assistant panel with receipts, nudges act…

    JavaScript

  4. Singleton Singleton Public

    Provably-fair, no-oversell allocation of scarce slots, on Amazon Aurora DSQL.

    TypeScript

  5. underVault underVault Public

    One shared dungeon. One candle a day. You get a single life, and the dungeon's rules are hidden from you. The only way to learn them is together: whatever one player discovers and banks into the Co…

    TypeScript

  6. datahub-project/datahub datahub-project/datahub Public

    The Context Platform for your Data and AI Stack

    Python 13k 3.8k