# Bisher — AI engineer (full text) > Bisher builds AI systems that take on a task, use software tools to finish it, > and check their own work. This file is the complete plain-text content of https://bishertello.com/, expanded in a single document for AI agents with a larger context budget. A condensed version is at https://bishertello.com/llms.txt. ## Overview Bisher is an AI engineer. He builds AI systems that take on a task, use software tools to finish it, and check their own work. The site is organised around four kinds of work — Agents, Models, Delivery, and Product — the skills behind them, and the topics he returns to. Currently working on: - Sending each request to the right agent with a small, fast model. - Evaluating agents, so a change can be checked before it ships. - Fine-tuning language and audio models for a specialized task. Kinds (used as colour keys throughout the site): - Agents — orange. - Models — green. - Delivery — blue. - Product — white. ## Signal path How the systems Bisher builds tend to flow, in order: 01 Trigger — a schedule or an event starts the work, so nobody has to ask. 02 Route (in progress) — a small, fast model picks the right agent. 03 Act — the agent uses tools to do the work. 04 Check — a review step looks at the result before it moves on. 05 Deliver — the result reaches a person or another system. ## Routing sketch An illustration of how a request is routed to the right kind of system. Three example requests, and the kind of system that takes each: - "Turn a stack of scanned documents into data." → Models → Read documents. A vision-language model reads each page and hands back structured data, whatever the document type. - "Write tomorrow's post, and have it checked." → Agents → Write and review. One agent drafts, another reviews, and the cycle starts again the next day without anyone asking. - "Show the team whether the agent got better." → Delivery → Test and evaluate. The agent runs against fixed checks, so a change is measured before it ships. The signal path and routing sketch are illustrative; nothing is sent anywhere. ## Can do What Bisher can build, and help decide to build, grouped by kind. Agents — software that acts, not only answers: - Start on their own: agents that begin work on a schedule or when something happens, so nobody has to ask. - Write and review: agents that draft content and agents that review it, on a daily cycle. - Route requests (in progress): sending each request to the right agent with a small, fast model. - Build workflows: an agent that creates business-process diagrams, with tools it can call to assemble the workflow. - Serve many customers: helped design a platform where many customers run their own agents, kept apart. Models — models that read, search, and listen: - Read documents: a modular system that turns documents into structured data with vision-language models, across document types and in batches. - Find answers: worked on the retrieval and search behind question-answering products. - Train and measure: fine-tuned and benchmarked language and audio models for specialized tasks, including distributed training. Delivery — getting it to people, and keeping it honest: - Ship the interface: dashboards, internal tools, and web apps for AI services, backend included. - Run on your own machine: tuned an agent to run entirely on a local computer, text recognition included, and to respond sooner. - Test and evaluate (in progress): built the tools a team uses to test and develop agents. Now building a pipeline to evaluate them. Product — the stage before building, deciding what to make and what to leave out: - Find the real problem: turning a rough idea into a specific problem — who has it, how they cope today, and what better would look like. - Shape the first version: deciding what the first version does, and what it leaves out. - Pick what to test first: choosing the cheapest thing that shows whether an idea works, before the whole product is built. - Fit it to what AI can do: matching an idea to what current models do well, what it will cost to run, and where they are heading. ## Good at Each rests on work Bisher has done. - Keeping cost in view: treats what an AI service costs to run as part of the design, not an afterthought. - Checking before trusting: builds the tests and reviews that sit around an agent, not only the agent. - Helping a team do its best work: coordinates the work, coaches people, and keeps a small team pointed the same way. - Design sense: cares how a product looks and reads, not only how it works. ## Likes What Bisher keeps coming back to. - Token economics — What thinking costs. Every model call costs money and time. He likes working out where that spend pays off and where it does not. - Agent harnesses — The part around the model. A model alone does little. The tools, memory, limits, and checks around it decide whether it is useful. That is the part he likes building. - AI development trajectory — Where this is going. He follows where model ability is heading and what it changes for the products built on top. - Product ownership — Owning the whole thing. He likes carrying a product from the first idea to the person using it. - Startups — Early companies. Small teams, short feedback loops, and a first version in front of real people early. - Benchmarks — Problems with a scoreboard. A clear score and a clock keep a problem honest. He likes work that can be measured that way. ## Contact Have something that should run on its own, or an idea to shape? Write to him. The email and profile are not published in this file and are not present in the site's HTML. The site shows them only after a Cloudflare Turnstile bot check, to prevent harvesting. To get in touch, use the contact section at https://bishertello.com/#contact.