Skip to content
View mallyma1's full-sized avatar
:bowtie:
Building
:bowtie:
Building

Block or report mallyma1

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
mallyma1/README.md

Mally Kisubi

Operations and delivery lead in London. I run client delivery for a data vertical and build the AI tooling it runs on.

Now. Regional Operations Team Lead at Chainlabs. I manage three client delivery leads and a 43-person contributor network across 25+ countries, own per-client SLAs, and write the client reports myself. I also build most of the vertical's tooling and ship it through a reviewed pull-request workflow.

What I've built there (employer code, so it stays private)

  • A multi-provider LLM gateway with automatic fallback, a cost ledger on every call and a two-tier budget kill switch that holds until a person clears it
  • Hybrid-search RAG, pgvector plus Postgres full-text search merged by reciprocal rank fusion, over the team's own documentation and split by audience
  • Evaluation sets built from human-verified quality-control results, and a regression gate that blocks a prompt change if the pass rate drops
  • The payout pipeline and the Telegram ops bot the contributor network works through every day

Public work

Repo What it shows
hybrid-search-measured My team's hybrid-search pattern rebuilt on a public finance benchmark (BEIR FiQA, 648 queries): rank fusion lifted nDCG@10 from 0.25 to 0.37 over keyword search, but vectors alone scored 0.40, so fusion has to be measured per corpus
ops-case-studies Operations decisions written up one page each: context, problem, what I did, result, evidence and limits
doomtax A solo ETHGlobal hackathon build (2026) with a "what didn't ship" file and a log of how AI tools were used. Live at doomtax.vercel.app
commitolab Commito, a commitment-tracking prototype built on Replit with its AI agent from written specifications, then cleaned and documented. Browser demo at commito-demo.vercel.app

Before this. Operations Manager at the Starknet Foundation, where I ran grants operations for 500+ organisations, designed the Foundation's first AML/KYB framework and cut onboarding and reporting cycle times by 33%. Before that, the first UK operations hire at Parity Technologies, where I set up the London office. Finance-trained, with a BSc from the University of York, part-CIMA, PRINCE2 and AgilePM.

mally-cv.vercel.app · linkedin.com/in/malcolm-k · kisubimalcolm@gmail.com

Popular repositories Loading

  1. doomtax doomtax Public

    Commitment device that stakes money on a focus session and streams the forfeit to charity. ETHGlobal Lisbon 2026.

    TypeScript

  2. mallyma1 mallyma1 Public

    GitHub profile README

  3. hybrid-search-measured hybrid-search-measured Public

    Hybrid search (BM25 + dense + rank fusion) measured on a public finance benchmark

    Python

  4. ops-case-studies ops-case-studies Public

    Operations decisions written up as one-page case studies

    Python

  5. commitolab commitolab Public

    Commito: a commitment-tracking prototype (Expo, React Native, Express, Postgres) with a browser demo

    TypeScript