yenklabs.com

Personal R&D lab. I'm working on one question: how do AI systems get the right context to make better decisions in real-world work? Experiments, notes, failures, and things I've changed my mind about.

github / linkedin / huggingface / about me
↑ contents

// Current Focus

The question I keep coming back to: how do AI systems get the right context to make better decisions in real-world work? Everything below is active research and development — none of it is solved, and none of it is productionized.

  1. Context assembly — selecting what actually matters for the task instead of dumping everything into a prompt.
  2. Context memory — preserving useful decisions, corrections, unresolved questions, and changing circumstances over time.
  3. Context boundaries — controlling what context an agent can access, and preventing inappropriate context leakage.
  4. Context validation — determining whether context is current, complete, conflicting, or insufficient before acting.
  5. Context learning — learning from human decisions and corrections without treating every previous decision as universal truth.

// Open Questions

Questions I'm actively working through, not ones I've answered:

// Follow the Context

I'm documenting what we're learning about context, memory, boundaries, and decision-making as AI moves from answering questions to doing real work.

Notes from the workbench. No AI news roundup.

A newsletter is starting soon. In the meantime, the work is here: GammaLex · LinkedIn · GitHub

// How This Connects

YenkLabs                    explore, question, prototype, write
    │
    ▼
Context Intelligence        the shared technical direction
    │
    ▼
GammaLex                    where promising ideas become real systems
    │
 ┌──────────┴──────────┐
 ▼                     ▼
Dali                  INOO
legal — matter        healthcare — care
context                context

Some experiments stay experiments. The ones that survive contact with real workflows, I take further through GammaLex, where we're building context agents for high-consequence professional work — that's where Dali and INOO come from.

Different work. Different rules. Same underlying challenge — AI is only as useful as the context it can understand, carry forward, and use responsibly.

This is a conceptual progression — research → shared direction → company → applications — not a software architecture diagram. There isn't one AI agent doing legal and healthcare; it's shared context capabilities underneath domain-specific adapters, workflows, permissions, and professional boundaries.

// Recent Activity

Most of this is the earlier verification thread that led here — dated, in order, unfiltered.

// The Forge

Dali and INOO are the GammaLex applications testing the direction above (see How This Connects). Everything else here is personal experimentation — no company attached.

dali → Legal

Legal intelligence built around the context of a matter, not isolated prompts. Current work explores how documents, claims, evidence, professional decisions, and prior work stay connected as a matter evolves.

inoo → Healthcare

Clinical intelligence built around the context of care, not a single encounter. INOO explores how conversations, clinical history, provider decisions, corrections, and changing patient context can stay connected while keeping the clinician in control.

yenk-inference-bench → Python

planned — Benchmarking local and hosted language models across latency, throughput, context utilization, and verification-oriented workloads.

micro-transformer-scratch → C++ / CUDA

planned — A transformer built from first principles to explore attention, training dynamics, memory movement, and inference mechanics without framework abstractions.

// Public Assets

What shipped from the earlier verification thread — code, datasets, specs:

Ecosystem mirrored on Hugging Face: yenklabs

// Platform Stack

What surprised me: catching a fabricated citation wasn't the hard part. The hard part was that the "why" kept receding — why did the model answer confidently with nothing behind it? Chasing that question rewrote the plan:

Why did AI answers fail? → Can we verify the evidence? → What information did the system actually use? → What information should it have used? → How should useful context persist and change over time? → What context should an agent be allowed to access? → How do we know its context is sufficient before acting? → Context Intelligence.

The stack below is that earlier leg of the journey — verification, evidence, benchmarks. Not the lab's current identity, but how I got here.

Open Evidence Infrastructure

Public corpora, EPS, Hugging Face datasets, and open-source verification engine.

Evidence Corpus

Portable evidence records — web browser · open-evidence-corpus on Hugging Face.

Verification Engine

Taxonomy, benchmarks, replay determinism, cryptographic lineage — Dali.

Enterprise Platform

Commercial evidence layer for regulated workflows — forthcoming.

// Investigations & Notes

Where the context problem first showed up as a concrete failure, starting with Mata v. Avianca: