Research Assistant · Skene Lab · London

Mrinal Joshi.

Neurogenomics, and the cool stuff I build between experiments.

I'm a Research Assistant in the Skene Lab at the UK Dementia Research Institute, Imperial College London. At the bench I work on microglia and prime editing. Off it, I build the software and AI agents the science needs. Current projects include deRoute, FLAIR-AD and Hearth.

See the work Play the demos ↗ Say hello
Lab · Skene · UK DRI HPC · Imperial CX3 Code · Python · R · TS Agents · Claude · Gemini · OpenCode
IRF1 pharmgx scRNA score=85 clocks agent mrinal.is-a.dev ↑ you are here microglia
§ 01 · About

Bench, code, agents.

The science covers neurogenomics and computational biology. The software exists because the science needs it.

single-cell sequencing microglial genetics Alzheimer's disease
Role
Research Assistant
Lab
Skene Lab, UK DRI
Institution
Imperial College London
Education
MSc Genes, Drugs and Stem Cells, Imperial (2025)
BSc Molecular Genetics, First Class, King's College London
Focus
Transcription-factor binding in microglia (IRF1)
Methods
Prime editing, TIP-seq, scTIP-seq, scRNA-seq, iPSC microglia
Compute
Imperial CX3 HPC

IRF1 sits upstream of several Alzheimer's risk genes (ApoE4, TREM2, others) in microglia. The wet work FLAG-tags it in iPSC-derived microglia via prime editing, then maps binding with CETCh-seq and TIP-seq. The dry work integrates those maps with bulk and single-cell RNA-seq plus GWAS fine-mapping.

The software below started as lab problems. A few grew sideways into personal infrastructure, and one of them, deRoute, won its track at Encode London.

§ 02 · Work

Thirteen projects, most of them running.

The six below get the long version. The rest are underneath. Most have a playable version at m-joshi.ai.studio.

01
Winner · Encode London 2026
your device Email [EMAIL_1] relay can't read it sealed computer C1 C2 C3 C4 C5 C6 C7 seven checks, then send

deRoute

Agents Public Live

Private AI you can check, not just trust. Your browser hides personal details, then seals the message so only one verified sealed computer can read it, after seven hardware checks. Gemini plans the agent's steps from an outline with none of your words. Won the Agents That Ship track at Encode London 2026, built solo in a weekend.

7checks before send ≈67bits max to Gemini ~500tests 1weekend, solo
TypeScriptNext.jsGemini 4 ArgonMongoDB AtlasTinfoil enclavesCircle Arc
Live → GitHub → Play the game →
02
PE7 TJ-PE PA ≤40 bp ≤500 bp >1 kb FLAG three modes, sized to insert

pegRNA design platform

Research Tools Live Private repo

The lab's prime-editing design platform for C-terminal protein tagging, and I'm its primary author. It picks the editing mode by insert size, then walks candidates through cheap sequence checks first and heavy models last, failing loudly if nothing passes.

PE7≤ 40 bp TJ-PE≤ 500 bp PA> 1 kb 880+tests
PythonFastAPIPRIDICT 2.0SpliceAIAlphaFold DBBoltz-2Imperial HPC
Live designer → Play → Private repo
03
IRF1

flair-ad

Research Public Live

Two-page companion site for the lab's IRF1 work: a public explainer and a scientist-facing data dashboard.

2pages 10+D3 charts 0build steps
D3.jsGSAPFirebaseNetlifyPlaywright
Live → GitHub → Play →
04
pharmgx gwas scRNA clocks ukb reproducibility bundle ✦ 24

ClawBio

Agents Tools Public · MIT

Local-first bioinformatics skill library for AI agents, where every run ships a reproducibility bundle. I've had four pull requests merged upstream: the skill-builder, which scaffolds new skills from a spec file, and three analysis skills.

4merged PRs 24skills 9databases 8k+Galaxy tools
PythonOpenClawCPICscVIPyAging
GitHub → Play →
05
rightmove foxtons get-living 0 100 score 85 08:00 UTC

hearth-ai

Agents Running Private repo

Agentic London rentals search. Crawl, hard-filter, tier-score with Claude vision, alert.

16sources 3-tierscoring ≥85instant alert 08:00 UTCdigest
FastAPISQLite WALCloudflare WorkersClaude Haiku / Sonnet 4.5
Site → Play → Private repo
06
T212 IG OAI Azure OpenRtr risk gate model decides → → risk filters 72h max hold · no shorts DRY-RUN

automated-trading

Agents Public Demo

Configurable CFD bot. Pluggable broker and model layers, with a deterministic risk layer in front of every order.

2brokers 3model providers 72hmax hold Dry-rundefault
PythonOpenAI ResponsesTrading212 / IGyfinance
GitHub → Play →

More projects

Seven more, from playable explainers to the plumbing behind my agents.

Portfolio · Live

Mrinal's AI Studio

Nine of my projects as playable mini-games, each followed by an honest write-up of what the real thing does.

React 19ViteTailwindGoogle AI Studio
Play →Private repo
Journal · Live

Positive Control

A builder's journal and hub for agentic biology, focused on checking what AI agents produce before anyone trusts it at the bench.

Next.jsTypeScriptVercel
Read →
Explainer · Live

A day at the bench

A playable day at my bench: tag a gene, pipette, look at my microscope images, then read a simulated qPCR run.

Next.jsTypeScriptVercel
Play →
Research · Hackathon

TF compensation-risk engine

Built with the five-person tractor-farma team at the Building an AI Scientist hackathon, co-hosted by Anthropic. Ten evidence axes feed ranked risk flags, and when my own metric failed its control, I reported the negative result.

PythonGitHub Actions560+ tests
Play →Private repo
Open source

OmniRoute

An open-source AI gateway with 80k+ stars: one endpoint for 160+ providers, with automatic fallback. I run my own fork and had Claude Messages API fixes merged upstream.

TypeScriptSelf-hostedTailscale
GitHub →Play →
Tools · Open source

OpenCode setup

My coding agent, with a memory that survives between sessions. Notes become rules only after they repeat, and every writer takes a lock.

OpenCodeClaudeGit hooks
GitHub →Play →
Agents · Running

Tutoring agents

Scheduled Claude agents for my A-level tutoring practice. They classify, draft and report, behind 769 offline tests, a dry-run mode and a halt flag.

TypeScriptClaudePlaywrightMCP
Tutoring site →Private repo
§ 03 · Stack

Day to day.

Tools in active use across the projects above.

Wet lab

iPSC microglia
Prime editing
TIP-seq · scTIP-seq
scRNA-seq

Code

Python · R · TS
FastAPI
Next.js · React
D3.js · GSAP

Agents

Claude Code
{} OpenCode
Claude · Gemini APIs
MCP

Infra

Imperial CX3
Docker
MongoDB · SQLite
workers Cloudflare Workers
§ 04 · Now

Current focus.

As of November 2026.

  1. Preparing my talk for the Festival of Genomics & Biodata, Agentic AI for Life Sciences: Build Your Own AI Research Assistant, on 27 January 2027.
  2. Running the pegRNA designer end to end on the web, with live docking behind it.
  3. Writing up Positive Control, my builder's journal for agentic biology.
  4. FLAG-tagging transcription factors with prime editing in iPSC-derived microglia.