Retrieval grounding, evaluation, and deployment built around your own information. $3,999/mo retainer, one senior engineer, no hand-offs.
- Scope
- Build
- Evaluate
monthly retainer
Who this is for
You have a product, a process, or an idea that needs a large language model to read, reason over, or generate text against information the model was never trained on. A chat-window prototype got you excited. It will not survive real customers asking edge-case questions. I build the model layer myself: retrieval grounding, evaluation, and the connections back to what you already run.
The pain today
- A demo built in a chat window answers well until a real customer asks something specific
- Generic AI tools guess at your policies instead of reading them
- No way to tell whether a prompt change made answers better or worse
- API usage climbs every month with no visibility into which calls earn their cost
- The one person who understands the prompt setup is the only one who can change it
The outcome you get
- A model layer grounded in your own information, not general training data
- An evaluation routine that catches a regression before a customer does
- Retrieval and prompt design written down, not held in one person's memory
- Deployment on your own infrastructure, so the model layer stays yours
- A monthly view of usage and cost so spending has a reason attached to it
The three parts that make this actually work
Calling an API from a chat window is not development. Large language model development is the discipline around that call: deciding what information the model sees before it answers, testing whether its answers hold up under real questions, and deploying the result somewhere it can run reliably.
Three parts make up the practice. Retrieval grounding decides what the model reads before it responds, so answers come from your information rather than a guess. Evaluation measures whether a change to that setup made results better or worse, instead of relying on a feeling. Deployment puts the whole thing on infrastructure you control, with logging and cost visibility built in from day one.
I work with OpenAI and Claude, choosing between them based on what your case actually needs rather than a default preference.
A model layer grounded in your own information, not general training data
Grounding answers in your own information
A model with no grounding answers from whatever it learned during training, which is a poor match for a specific product, a specific policy, or a specific customer base. Grounding fixes that by retrieving relevant information first and handing it to the model as context before it responds.
I built this pattern for Instill, my self-initiated AI product launched in the first quarter of 2026. Instill is a knowledge base reached through the Model Context Protocol (MCP), an open standard for connecting language models to real data instead of training-time knowledge. It now runs with 30+ active users, 1,000+ skills saved, and 45+ projects powered. Running a live product through real usage exposed the failure cases a demo never surfaces: sparse results on unusual questions, context that runs long, and answers that sound confident while being wrong. Every custom build I take on now carries those lessons forward.
30+: Active users.
How I know a change actually helped
Prompt tweaking without a way to measure the result is guesswork dressed up as progress. Before I change a prompt, a retrieval setting, or a model choice, I write down what a correct answer looks like for a representative set of real questions. After the change, I run the same set again and compare.
This is not a one-time step. Every update to your information runs back through the same check before it reaches production. If a change lowers the pass rate, it does not go live until the reason is found. This is the same discipline I apply to any production system: know what changed, and know whether it changed things for better or worse, before it reaches a real customer.
Connecting it to what you already run
A grounded, evaluated model is only useful once it reaches the tools you and your customers already use: writing to a database, calling an existing API, or posting into a support tool, not living in an isolated chat window.
I deliver on infrastructure you own: AWS or Vercel, PostgreSQL or MongoDB, a Next.js or Node.js front door depending on where your product already lives. Work Made for Hire terms mean the code, the infrastructure setup, and the documentation belong to you from the day you pay, not to me.
Pricing, delivery, and what you keep
LLM development runs under the AI Development retainer at $3,999 a month. Work lands in cycles of two to four days rather than a single delivery at the end of a long build, so you see progress every week, not just at a deadline.
Communication is daily and async, with a response inside 24 hours. The engagement carries a 14-day money-back guarantee: if it is not working for you in the first two weeks, you get a full refund. After that, cancel anytime. Everything built is yours under Work Made for Hire terms from the moment you pay.
Recent proof
A comparable engagement, delivered and documented.
An AI knowledge base your whole team uses via MCP
A personal library for Skills, Agents, and Rules, built once, used across Claude, Cursor, and any MCP-compatible AI tool.
Read the case studyRelated services
The same work, framed for a different situation.
Frequently asked questions
The questions prospects ask before they book.
ChatGPT and similar consumer tools answer from general training data with no memory of your specific information. LLM development means building the layer around a model, retrieval grounding, evaluation, deployment, so the answers come from what you actually know and hold up under real use. The underlying model (OpenAI or Claude) is often the same. The difference is everything built around it.
I work with both and choose based on what the use case needs rather than a default. Some tasks favor one model's instruction-following, others favor cost or context length. I do not commit to a model before testing it against your actual questions.
Work lands in cycles of two to four days rather than in one large delivery. You see something running early, and it gets refined from there. Every update goes through the same evaluation check before it reaches production.
The engagement carries a 14-day money-back guarantee. If it is not working for you in the first two weeks, you get a full refund. After that, you can cancel anytime; there is no long-term lock-in.
You do, in full, from the moment you pay. Work Made for Hire terms cover the code, the infrastructure setup, and the documentation. Nothing stays with me after the engagement ends.
No. Readers of this page range from someone with only an idea through to an established business. What matters is having real information the model should answer from, whether that is a handful of documents or a full production database.