We fix, build and verify LLM systems for US businesses.
Fixed price. Dated deliverable. A paid first step that is credited to the build. Assistants that cite their sources, migrations off retiring models, lower token bills, and agents that are tested before they ship.
- Delaware corporation
- Fixed prices
- First step credited to the build
- Your data stays in your accounts
- NDA within one business day
Each of these is a one- or two-week job with a fixed price.
These are the problems US owners and teams are writing about this fall. Pick the one closest to yours.
Our assistant makes things up, and we upload files to it by hand every week.
Assistant that cites or refusesOur model is being shut down and we don't know every place it's called.
Model-retirement migration sprintThe token bill grew three times. Usage didn't.
LLM cost and latency reductionWe built it ourselves with Claude. We need someone to tell us if it’s right.
Architecture and readiness reviewOur custom GPTs retire on December 11 and the actions don't carry over.
Custom GPT migrationThe agent broke after an update and nobody noticed for a week.
Agent reliability and evals harnessThe next four shutdowns land across the holidays.
Teams that fixed the October 23 breakage by moving to a newer model often landed on one that shuts down next. The calendar lists every date, provider and code change we track.
-
Nov 30 2026
- Claude Sonnet 4.5 retires
- OpenAI Evals, Agent Builder and the Prompts API shut down
-
Dec 11 2026
- GPT-5 and o3 models shut down
- ChatGPT custom GPTs retire
-
Jan 6 2027
- tts-1 and gpt-4o-mini-tts shut down
- New fine-tuning jobs end
-
Jan 14 2027
- GPT-4o, GPT-4o mini, GPT-4.1 and their fine-tunes shut down
Nine fixed-scope packages. Every one starts with a small paid step.
Prices are published so you can decide before a call. The first step is credited to the build if you go ahead within 30 days.
Model-retirement migration sprint
For teams whose code calls models that shut down between November and January.
- Price
- $2,500
- Time
- 5 business days
Start with Migration audit, $1,200
Custom GPT migration
For ChatGPT Business and Enterprise workspaces with custom GPTs people rely on.
- Price
- $1,200
- Time
- 3 business days
Start with Inventory and migration plan, $400
Assistant that cites or refuses
For teams whose assistant answers from internal documents and gets it wrong.
- Price
- $3,000
- Time
- 10 business days
Start with Single-source starter, $1,200
Document extraction pilot
For firms re-typing invoices, leases, tax forms, rate confirmations or rent rolls.
- Price
- $1,500–$2,800
- Time
- 5–10 business days
Start with Proof on 10 documents, $750
LLM cost and latency reduction
For products and teams whose model spend is rising faster than usage.
- Price
- $1,500 audit
- Time
- 1 week audit; 1–2 weeks implementation
Start with Cost audit, $1,500
Architecture and readiness review
For founders and teams who want an independent review before or after building.
- Price
- $1,200–$1,500
- Time
- 3–5 business days
Start with the review itself
Agent reliability and evals harness
For teams running agents or LLM features without automated tests, or losing OpenAI Evals on November 30.
- Price
- $2,000–$3,500
- Time
- 1 week
Start with Baseline and plan, $1,200
Workflow automation (n8n, Claude, MCP)
For owners connecting email, CRM, documents and AI, and finding the joins break.
- Price
- $2,200
- Time
- 5–10 business days
Start with Two-day repair, $950
Financial data assistant (text-to-SQL)
For finance teams, advisers and fintechs who want answers from their own database.
- Price
- $4,500
- Time
- 10–15 business days
Start with One-week prototype, $1,500
Systems our founder has already built
Real numbers from production systems, with the role stated on each case study. Client names are withheld where the work was confidential.
20,000
users on an analytics product answering in under two minutes, at ~10% of equivalent AWS cost
Case study1 in 10
AI-generated items caught by a deterministic quality gate across ~50,000 items
Case study95%
less report review time with answers from SEC filings, each traceable to its source
Case studyOpen source
UnvibeCode maps what AI-written code actually does, with evidence-backed risk findings
Case studyFive steps, written down before we start
50% to start and 50% on acceptance. Acceptance criteria are agreed in writing first. The founder does the work himself; nothing is handed off.
- 01
20-minute call
You describe what you have and what is going wrong.
- 02
Written scope
Acceptance criteria and a fixed price, within one business day.
- 03
NDA and access
We sign your NDA and work inside your accounts.
- 04
Build
A short written note at the end of every working day.
- 05
Handover
Evaluation set, runbook, recorded walkthrough and all code.
Built around the rules your industry already follows
Accounting, tax and bookkeeping firms
Document intake, email triage and a cited procedures assistant, built around IRC §7216.
IRC §7216 consent before return data reaches a vendor
Property managers
Maintenance triage, tenant email and owner reporting, with people kept on money and deadline replies.
Fair Housing Act: no screening decisions by a model
Commercial real estate
A deal assistant that never invents a property, and rent-roll and T-12 extraction with the math in code.
Every figure computed in code and linked to its page
Solo and small law firms
Intake automation and a cited assistant over your own library, designed around ABA Formal Opinion 512.
ABA Formal Opinion 512 on client confidentiality
AI startups and product teams
Model migrations, cost reduction, evals and an independent review, for teams shipping on model APIs.
Model names in one config, so the next notice is small
We review other people's claims for a living
The same discipline that reviews manuscripts for journals reviews your system: what is claimed, what the evidence shows, and what would break it.
- Peer-reviewed, 2026
First-author paper in International Review of Economics & Finance (Elsevier).
- Preprint, 2026
Evidence-bound factual repair in retrieval-augmented LLM answers.
- Peer review
Reviewer for Elsevier and Springer Nature journals in applied AI and computing.
- Program committee
Program-committee reviewer, NeurIPS 2026 Verify-Agents workshop.
Tell us what you have and what is going wrong.
A 20-minute call, then a written scope with acceptance criteria and a fixed price within one business day. If it isn't a fit, we'll say so on the call.
Or email hello@destello.us or call (628) 202-5322. We reply within one business day.