Build on research.
Innovation

Nairobi research and innovation firm. Field research turned into tools, datasets, and papers.

Research & Outputs

Open source products and published papers

Tools, datasets, and working papers from the lab.

Papers

ai ml 

The Asymmetric Turn: Inside DeepSeek V4.1 Flash and Why It Wins

Tawakkul Labs19 Sept 202612 min read

DeepSeek V4.1 Flash replaces the uniform decoder with a 40 layer Causal Encoder Decoder: 20 encoder layers read the prompt with 8B active parameters and 20 decoder layers write with 16B. Alongside a compressed sparse attention scheme and a much smaller KV cache, it changes what long context agent work costs. This paper examines the architecture, the reported benchmarks, the API economics, and the honest caveats, and asks what a vendor reported win on agent tasks means for teams building on African hardware budgets.

Read paperPDF
ai ml 

MoE at the Edge: Offloading and Small Mixture of Experts Models for Constrained Hardware

Tawakkul Labs17 Sept 20268 min read

A technical survey of two parallel lines of work: expert offloading, which moves mixture of experts weights between memory tiers so large models run on constrained devices, and small mixture of experts models, which compress the architecture itself to sub billion scale. It covers the routing consistency caveat, predictive offloading, adaptive precision, expert merging, the on device MoE scaling law, and the honest open question of whether MoE actually helps on consumer hardware, all examined for what it means on African hardware budgets.

Read paperPDF
ai ml 

State of AI Architecture 2026

1 Aug 20265 min read

A synthesis of four architectural shifts that defined 2026: mixture of experts becoming the frontier standard, the return of linear state space models with Mamba 3, the formalization of agentic retrieval augmented generation, and the production lessons of multi agent orchestration. Each trend is examined for what it means on African hardware budgets.

Read paperPDF
ai ml 

Local AI Architecture for Rural and Informal Sectors

1 Aug 20269 min read

An audit of what open source AI can actually do on low end devices with no internet: the hardware tiers, the quantization and runtime stack that makes small models viable, the models that fit each tier, and proven use cases from East Africa. Ends with tiered advice for building local AI systems for rural and informal sectors.

Read paperPDF

Products

mcp security 

GRIM

v0.5.0MITMCP

Security audit for AI agents. GRIM finds code, dependency, exposure, secrets, and active compromise gaps across any stack, with drift detection and an SBOM.

Read moreGitHub ↗
mcp code quality 

Harmony

v0.1.5MITMCP

Repo rules enforcement for AI agents. Harmony checks diffs against the repository rule set and blocks violations before they land.

Read moreGitHub ↗
mcp memory 

Logbook

v0.1.3MITMCP

Project memory across sessions for AI agents. A queryable decision log that survives between sessions.

Read moreGitHub ↗
mcp integrity 

Reference

v0.1.6MITMCP

Reference integrity for AI agents. Scans a repository for every reference and verifies each one resolves.

Read moreGitHub ↗
mcp release 

Release Gate

v0.1.3MITMCP

Release intelligence for AI agents. Analyze changes, detect breaking updates, and recommend the next semantic version.

Read moreGitHub ↗
mcp deployment 

Targetly

v0.1.4MITMCP

Deployment truth verification for AI agents. Targetly answers whether the live site is what you just built.

Read moreGitHub ↗
mcp tokens 

Token Gate

v0.1.3MITMCP

Token and context budget for AI agents. Track spend, estimate the cost of a read, and catch the burn before it happens.

Read moreGitHub ↗
flutter dart 

Heides Lens

v1.0.0MITMCP

The screen for HEIDES. A desktop viewer and query tool over the HEIDES code nervous system.

Read moreGitHub ↗
rust cli 

HEIDES

v0.14.4MITMCP

The code nervous system. HEIDES gives any LLM instant, deterministic understanding of a codebase, so agents stop reading files and start knowing them.

Read moreGitHub ↗
./research --all
The group

Research. Engineered. Shipped.

Tawakkul Labs establishes the evidence base. Taw turns it into the software, AI, and digital systems that move East African businesses forward.

Tawakkul Labs

Tawakkul Labs

Research core

We design and conduct field research across East Africa, then publish the findings, datasets, and open source tools that inform serious decision making.

Explore the research
TAW

Taw

Innovation & development

We engineer the software platforms, AI automations, and digital marketing systems that take research from the lab to production, at speed and at scale.

Visit Taw
About us

The team behind the research

Tawakkul Labs is a Nairobi based research and innovation firm. We conduct field research across East Africa and turn our findings into published work, open source tools, and datasets.

We exist to close the gap between academic research and the builders who need it most, so the next generation of African tools is built on local evidence.

Development is spearheaded by Taw, our innovation and engineering arm. It turns the lab's research into software, AI automations, and digital systems that scale.

Partner with us

Research first, tools second

We build open infrastructure for East African innovation. No proprietary lock in, no black boxes.

  • ✓ Open source research with reproducible code
  • ✓ Open source datasets and tools
  • ✓ Collaborations with Nairobi builders
Contact

Collaborate with us

Research partners, implementers, funders, and open source contributors welcome.

0794 751 252
0737 190 114