AI | Assah Bismark

Oct 14, 2026

I Started a Quantitative Research Firm

Why I started Aveling Research.

It still feels strange to call it a company, but it is one. I started Aveling Research, a quantitative research firm founded in Cameroon. why Aveling exists My background is in software engineering and AI engineering. Aveling grew out of applying that background to systematic trading research. I beg...More ›

Sep 1, 2026

The Chat App Is the Last Thing That Matters

Teams, Slack, Discord, a custom GUI. People keep asking where agents should live. The control plane is a different thing, and it isn't a surface.

People keep asking which chat app agents are going to win. Teams, Slack, Discord, or a custom GUI. A team moving to agents picks a surface before it has anything to put behind it. I think the question is backward. The chat app is just where the dispatch message lands, and agents stopped being chat a...More ›

Jun 17, 2026

Making Snif Fast Enough to Run on Every PR

How we removed redundant LLM calls, eliminated hard caps, added code embeddings, and built evidence verification into the review pipeline.

A repository with 5,000 symbols. One line changes. Snif re-summarized all 5,000 symbols, re-embedded all 5,000 summaries, then ran retrieval on the full set. Roughly 1,000 LLM calls and 5,000 embedding calls to review a single-line diff. 99% of those symbols hadn't changed since the last index. Thei...More ›

Jun 1, 2026

A Sandbox Is Not an Agent Runtime

A sandbox gives an agent a machine. The runtime gives it identity, state, approvals, logs, and recovery.

Code is still one of the central assets of a company. AI agents don't make it cheap. They make it easier to produce more of it, which means they also make it easier to produce more mistakes, more migrations, more half-finished branches, and more changes that someone has to understand. That is why ar...More ›

May 14, 2026

Spikes Are for Technical Bets

Storage needed numbers. The reviewer needed a phase constraint.

Spikes are for technical bets. Snif had two architecture questions before the main implementation: where repository context should live, and how the reviewer should run. Snif is an AI code review tool. It analyzes a pull request, retrieves the relevant context from the repository, and runs a structu...More ›

Apr 12, 2026

Stabilize Claude Code for Open-Weight Models

A three-layer proxy stack to keep Claude Code stable when routing through LiteLLM to open-weight models.

Running Claude Code with open-weight models like DeepSeek or Qwen through a LiteLLM proxy works. Until it doesn't. The port changes every time LiteLLM restarts because it picks a random dynamic port. The request payload grows until backends reject it with 400 errors. Usage stats come back null and t...More ›

Apr 10, 2026

Factory Droid with Any OpenAI-Compatible Endpoint

Use LiteLLM as a local proxy to run Factory droid with any OpenAI-compatible model endpoint.

Factory droid supports custom models through BYOK (Bring Your Own Key), but environment variables don't work for the interactive CLI or desktop app. The droid daemon has its own API client and ignores and env vars. The working path is adding your model to the array in and pointing it at a local ...More ›

Apr 1, 2026

Hacker News Is High Friction

HN is the best signal source in tech. Reading it shouldn't cost an hour.

Hacker News is where the best technical conversations happen. When a new database gets traction, the thread on HN is where the people who actually built competing systems show up and explain the tradeoffs. When a paper drops, the comments are often better than the paper. The signal is real and it's ...More ›

Apr 1, 2026

Context Is What Makes Code Review Work

AI code review is a systems problem, not a prompt problem. Here is how I built Snif to stay useful.

I spent some time researching every AI code review tool I could find. CodeRabbit, Greptile, Qodo merge, Copilot and lots of others, half a dozen smaller ones. The pattern I saw seems to be the same, teams install these tools, get excited with the feedback they provide, then disable them after some t...More ›

Mar 26, 2026

Using Claude Code with any model

Claude Code's toolset works with non-Anthropic models through a local proxy. Here's the setup, the bug, and the tradeoffs.

Claude Code has the best agentic coding toolset available right now. File editing with diffs, bash execution, grep, glob, MCP server integration, subagents, plan mode. It runs in your terminal, reads your entire codebase, and orchestrates multi-step changes autonomously. It only speaks Anthropic Mes...More ›

Mar 19, 2026

AI Shipped Faster Than Anyone Can Verify

AI systems ship at exponential speed. The governance and security infrastructure around them is still moving at human speed. That mismatch is the actual problem.

I've been tracking what's actually happening in the AI space day to day. Not product launches or funding rounds. The pain. What's breaking, what's failing, what practitioners are complaining about. The same five problems keep showing up. Agent overload. AI-generated code debt. OSS maintainer burnout...More ›

Mar 15, 2026

AI coding agents and the cloud provider gap

Desktop AI coding agents don't work with cloud provider auth. A local proxy fixes all of them at once.

The current generation of AI coding agents ships with a shared assumption: you have an API key from a model provider, and you paste it into a settings field. That works if you pay for a direct subscription. It does not work if you access models through a cloud provider like AWS Bedrock, Google Verte...More ›

Mar 10, 2026

Architecture Is the New Product

AI writes the code now. The architecture is what actually ships.

What do software teams actually produce? Most people say features. Ship features, hit deadlines, move the roadmap. For a long time that was close enough. But now AI can generate features. It can scaffold services, write CRUD endpoints, wire up frontends, produce tests. The code itself is approaching...More ›

Mar 4, 2026

Taste, Judgment, and the Thing AI Cannot Do

AI handles the logic. But logic was never the hard part.

AI has been around since the very beginning of computers. This is not some new phenomenon. Computing has steadily evolved from binary code and assembly language to the high-level languages we use today, all of it trying to bridge the gap between what a human wants and what a machine can execute. Pre...More ›

Mar 4, 2026

There Is No AI Thinking (And You Can't Outsource It)

The machines got faster. The real question is whether we got lazier.

There Is No AI Thinking (And You Can't Outsource It) We have officially entered the era of Inference-at-Scale. The marketing hype surrounding "Artificial Intelligence" has never been louder, bolstered by the deployment of NVIDIA's Rubin) architecture and the rise of Agentic Ecosystems. These systems...More ›

Jan 21, 2026

The Infinite Software Crisis

What Happens When AI Writes Faster Than We Can Think

The Infinite Software Crisis: What Happens When AI Writes Faster Than We Can Think There's been a lot of talk lately about what's being called the "Infinite Software Crisis". It's not a new observation - people have been warning about this for a while - but it's been circulating more in engineering ...More ›

Nov 1, 2025

Solving the Context Problem - A Local RAG System for Code

Building Local Semantic Code Search - 11ms Queries Across 7,620 Files

Solving the Context Problem: A Local RAG System for Code Yesterday I was using Kilocode to help refactor some authentication code in a 7,620-file Java codebase. The AI agent kept giving me generic advice because it couldn't see the actual implementation patterns in the project or some constraints. I...More ›