agentic engineering / production ai agents that ship
we ship autonomous agents inside your product. state, memory, tool use, guardrails, evals. live in weeks, in your repo, on your own claude account.
lock your architecture scoping (usd 4,750 / thb 156,400)
00 . the short answer
Architecture discovery is USD 4,750 (THB 156,400), a fixed fee credited to the build when you proceed. The build runs from USD 24,500 (THB 806,800) per phase, fixed price. Terms are 50/50 under USD 25k and 30/40/30 at USD 25k and above, paid by Stripe, Wise, or Thai bank transfer in THB. Every price is published on this page, which is more than any agency in Bangkok will show you before a call.
One to two weeks of discovery, then six to fourteen weeks of build. The written architecture spec and eval plan land before the first line of the agent loop is written. Build length tracks how many tools and integrations the agent touches. Launch comes with 90 days of care and two tuning passes against real production traffic.
AI agent engineering for defined product and operational tasks, with permissioned tools, evaluation cases, failure handling and human decision points. The delivery plan follows technical discovery.
01 . what it is
AI agent engineering for defined product and operational tasks, with permissioned tools, evaluation cases, failure handling and human decision points. The delivery plan follows technical discovery.
a chatbot is a text box wired to an api key. an agent is a system. it holds state across steps, decides which tool to call and when, retries when a call fails, and stops itself before it does something it should not. we build the second one. an autonomous support agent that reads a ticket, checks three internal systems and drafts the reply your reviewer approves in ten seconds instead of ten minutes. a document pipeline that ingests thai and english invoices, extracts line items against your own catalog, and writes structured rows to postgres with a confidence score attached. a research agent that runs six searches, cross-checks the sources and returns a memo instead of a wall of links. that is what ships here, and it ships running.
AI agent engineering for defined product and operational tasks, with permissioned tools, evaluation cases, failure handling and human decision points. The delivery plan follows technical discovery.
it ships into your codebase, not ours. github repo under your org from the first commit, supabase project under your billing, claude usage on your own anthropic console. you own every line on day one and you keep shipping it even if you never speak to us again. already running an internal tool on next.js and supabase? we add the agentic layer straight into that codebase. auto-triage on the support queue, auto-draft on a document type, auto-reconcile on the finance export. no rebuild, no platform tax, no bolt-on tool wrapped around your product.
AI agent engineering for defined product and operational tasks, with permissioned tools, evaluation cases, failure handling and human decision points. The delivery plan follows technical discovery.
the engagement opens with a paid technical discovery, never a free strategy call. it produces a written architecture spec, state model, tool inventory, memory strategy and failure modes, plus the eval plan with the exact test cases and pass thresholds your build is measured against. all of it lands before one line of the agent loop is written. usd 4,750, fixed, credited to the build. the document is yours whether or not you build with us. every price sits published on this page, which is more than any agency in this city will put in writing before a call.
02 . who it is for
illustrative scenario 01
Illustrative planning scenario. AI agent engineering for defined product and operational tasks, with permissioned tools, evaluation cases, failure handling and human decision points. The delivery plan follows technical discovery.
illustrative scenario 02
internal ops tool already running on next.js and supabase, three years old, thousands of rows moved by hand every week. kade wants zero rebuild: auto-triage on the intake queue and auto-draft on the weekly report, shipped straight into the codebase that already exists. the agent respects the rls policies and the audit log already in place. we work inside your repo, so nothing gets thrown away and nothing gets rented back to you.
illustrative scenario 03
multi-warehouse fulfilment across three marketplaces, shopee, lazada and a shopify own-site, with reconciliation done by two people in a spreadsheet every friday. aew gets a multi-step agent wired into the existing ops app: read the marketplace order feed, cross-check stock across warehouses, flag mismatches, draft the reconciliation entry, and escalate to a human the second a number looks wrong. friday afternoon goes back to the team.
03 . what you receive
01
architecture spec and eval plan. written during discovery: state model, tool inventory, memory strategy, short-term context against long-term retrieval, failure modes, and the exact eval cases and pass thresholds your build is measured against. it lands before the first line of the agent loop exists, and it is yours to keep.
02
agent implementation. the core agent loop shipped inside your next.js and supabase codebase: tool use wired through a typed, permissioned layer instead of raw api keys handed to a model, retries with backoff, guardrails that stop an out-of-scope action before it executes.
03
retrieval and memory layer. a rag pipeline over your product catalog, knowledge base or document set so the agent answers from your data instead of guessing from training data. short-term state for multi-step tasks, long-term memory where the workflow demands it.
04
observability and logging. every agent run traced: which tools were called, what the model saw, what it decided, what it output. a slack or line channel firing anomaly alerts during bangkok business hours, so you see a problem before a customer does.
05
documented eval suite. the test cases from the architecture spec wired into ci, so a prompt change, a model upgrade or a new edge case is caught before it ships, not after a customer complains.
06
handover documentation and post-launch tuning. a 20-page operating manual covering the agent architecture, the eval suite and the guardrail logic. 90 days of post-launch care: bug fix sla and two tuning passes measured against real production traffic.
04 . timeline
phase 0
1 to 2 weeks
paid technical discovery. stakeholder interviews, a full review of your existing codebase, state model design, tool inventory, memory strategy and failure-mode mapping. you walk out with the architecture spec and the eval plan, signed before a build day starts.
phase 1
weeks 1 to 3 of build
the state machine, the model-calling layer, the first tool integrations wired. weekly demo with the engineers who wrote the code and your named project lead. by the end of week 3 your agent completes a multi-step task end to end.
phase 2
weeks 4 to 6 of build
the full tool inventory wired, internal apis, database reads and writes, external services, with permission boundaries enforced per tool, retry and backoff logic, and the guardrails that kill an unauthorized action before it executes.
phase 3
weeks 7 to 9, plus 90 days
the eval suite runs against the full test set, observability goes live, production cutover, handover documentation delivered. then 90 days of post-launch care, tuning against real traffic instead of synthetic test cases.
05 . pricing
paid technical discovery: usd 4,750 (thb 156,400), fixed fee, credited to the build when you proceed. it covers stakeholder interviews, a full codebase review, the architecture spec and the eval plan, and you keep the document either way. build runs from usd 24,500 (thb 806,800) per phase, fixed price, the same discipline as our custom apps builds: 6 to 14 weeks end to end depending on how many tools and integrations your agent touches. thb shown at 32.93 to the dollar, bot reference rate at invoice date is what bills. usd or thb both accepted. terms: 50/50 for engagements under usd 25k, 30/40/30 for engagements usd 25k and above. stripe usd, wise usd, or thai bank transfer in thb. every number is on this page before you ever speak to us.
06 . proof
Selected work
Jobfin is a recruitment platform associated with Belgian federal finance recruitment. The published case describes DLVX delivery scope; it is not evidence of an AI deployment or a security certification.
recruitment platformpublished delivery case07 . questions
chatgpt or claude in a browser tab is a person typing a prompt and reading an answer. what we build is a system: the agent holds state across a multi-step task, calls your internal tools and apis through a permissioned layer, retries when a call fails, and passes an eval suite before and after it ships. the model call is the smallest part of the work. the state machine, the guardrails, the observability and the retrieval layer are the engineering, and none of them exist when someone opens a chat window.
guardrails are designed during discovery, never bolted on after an incident. every tool the agent can call is permissioned and scoped: an agent that drafts a refund cannot execute one without a human approval step unless the architecture spec authorizes exactly that. retries have backoff and a hard stop. every run is logged, so an anomaly surfaces in minutes instead of arriving as a customer complaint. the eval suite runs the agent against known edge cases before every deploy, failure cases included.
yes, and completely. github repo under your org from the first commit, supabase project under your billing, claude usage on your own anthropic console. the agent ships inside your codebase, never a tool you rent from us. if we vanished tomorrow, your engineers keep shipping it on monday.
yes, they stack. claudekick automates operational workflows on top of the tools you already run: your accounting software, your line oa, your crm. agentic engineering ships agent capability inside the product itself, a feature your customers and your team touch every day. run claudekick for the finance team's monthly close and a support agent from this service inside the product, both on one claude account, zero conflict.
a documented set of test cases your agent passes before it ships and before every future deploy: known inputs, expected behavior, pass or fail thresholds, edge cases and failure modes included. it is wired into ci, so a prompt change, a model version upgrade or a new tool integration gets checked automatically instead of waiting for someone to notice the agent got worse. this is the line between an agent you trust in production and a demo that worked once.
next.js and supabase as the application layer, the same stack as our custom apps builds. claude for the model layer, on anthropic's tool-use apis for the agent's actions. postgres with pgvector or a dedicated vector store when retrieval is part of the design. sentry and supabase logs for observability, ci-wired eval runs on every deploy. aws in the region you choose, or fly.io when latency or data residency decides it.
Architecture discovery is USD 4,750 (THB 156,400), a fixed fee credited to the build when you proceed. The build runs from USD 24,500 (THB 806,800) per phase, fixed price. Terms are 50/50 under USD 25k and 30/40/30 at USD 25k and above, paid by Stripe, Wise, or Thai bank transfer in THB. Every price is published on this page, which is more than any agency in Bangkok will show you before a call.
One to two weeks of discovery, then six to fourteen weeks of build. The written architecture spec and eval plan land before the first line of the agent loop is written. Build length tracks how many tools and integrations the agent touches. Launch comes with 90 days of care and two tuning passes against real production traffic.
AI agent engineering for defined product and operational tasks, with permissioned tools, evaluation cases, failure handling and human decision points. The delivery plan follows technical discovery.
An architecture spec and eval plan, the agent implementation with typed tool use and guardrails, a retrieval and memory layer, observability on every agent run, a documented eval suite wired into CI, a 20-page operating manual, and 90 days of post-launch tuning. All of it inside your own Next.js and Supabase codebase, owned by you from the first commit.
It depends who touches it. A Zapier or Make flow with a Claude node has no state, no memory, no eval suite, and no answer for an edge case at 3am. That is fine for one-off internal automation, which DLVX covers under AI integration. An agent is for the feature paying customers and staff use inside your product every day.
A chat window is a person typing a prompt and reading an answer. An agent holds state across a multi-step task, calls your internal tools and APIs through a permissioned layer, retries when a call fails, and passes an eval suite before and after every ship. The model call is the smallest part of the engineering.
Yes. The GitHub repo sits under your org from the first commit, the Supabase project is under your billing, and Claude usage runs on your own Anthropic console. The agent ships inside your codebase, never a tool you rent from DLVX, so if DLVX vanished tomorrow your engineers keep shipping it on Monday.
Guardrails are designed during discovery, never bolted on after an incident. Every tool the agent can call is scoped and permissioned, so an agent that drafts a refund cannot execute one without human approval unless the spec authorizes it. Retries have a hard stop, every run is logged, and the eval suite runs before every deploy.
07b . คำถามที่ถามบ่อย
จ้างทำ AI agent ราคาเท่าไหร่?
เริ่มที่ USD 4,750 (THB 156,400) สำหรับเฟสออกแบบสถาปัตยกรรม เป็นราคาเหมาจ่าย และหักเป็นเครดิตให้เมื่อทำต่อถึงขั้นพัฒนาจริง ส่วนงานพัฒนาเริ่มที่ USD 24,500 (THB 806,800) ต่อเฟส ราคาคงที่ แบ่งจ่าย 50/50 ถ้างานต่ำกว่า USD 25,000 และ 30/40/30 ถ้าตั้งแต่ USD 25,000 ขึ้นไป ชำระผ่าน Stripe, Wise หรือโอนเข้าบัญชีธนาคารไทยเป็นเงินบาทได้ ทุกราคาแสดงไว้บนหน้านี้หมดแล้ว ก่อนคุยกันด้วยซ้ำ
AI agent ต่างจาก chatbot ทั่วไปยังไง?
chatbot คือกล่องแชทที่ต่อ API ไว้เฉยๆ ส่วน agent เป็นระบบที่เก็บสถานะงานข้ามหลายขั้นตอน เลือกเรียกเครื่องมือภายในผ่านชั้นที่กำหนดสิทธิ์ไว้ชัดเจน ลองใหม่อัตโนมัติเมื่อเรียกไม่สำเร็จ และต้องผ่านชุดทดสอบ eval ก่อนขึ้นระบบจริงทุกครั้ง งานส่วนใหญ่จึงเป็นงานวิศวกรรมซอฟต์แวร์ระดับเดียวกับระบบชำระเงิน ไม่ใช่แค่การเขียน prompt
ทำ AI agent ใช้เวลากี่เดือน?
เฟสออกแบบสถาปัตยกรรมใช้เวลา 1-2 สัปดาห์ ได้เอกสาร architecture spec และแผน eval ก่อนเริ่มเขียนโค้ดจริง จากนั้นงานพัฒนาใช้เวลา 6-14 สัปดาห์ ขึ้นอยู่กับว่า agent ต้องเชื่อมต่อกับระบบและเครื่องมือภายในกี่ตัว หลังขึ้นระบบยังดูแลต่ออีก 90 วัน พร้อมปรับจูน 2 รอบจากการใช้งานจริง ไม่ใช่จากเคสทดสอบ
โค้ดที่จ้างทำ AI agent เป็นของใคร?
เป็นของคุณทั้งหมดตั้งแต่วันแรก โค้ดอยู่ใน GitHub repo ขององค์กรคุณตั้งแต่ commit แรก ฐานข้อมูลอยู่ใน Supabase project ที่จ่ายค่าบริการในนามบริษัทคุณเอง และการเรียกใช้ Claude ผูกกับบัญชี Anthropic ของคุณ agent ติดตั้งอยู่ในโค้ดเบสของคุณ ไม่ใช่เครื่องมือที่ต้องเช่าจาก DLVX ถ้าวันหนึ่งไม่ได้ทำงานกันต่อ ทีมพัฒนาของคุณรันเองได้ทันที
ธุรกิจแบบไหนควรลงทุนทำ AI agent จริงจัง?
DLVX วางแผนและพัฒนางานตามขอบเขตที่ตกลงร่วมกัน ตรวจสอบระบบและข้อมูลก่อนเริ่มงาน พร้อมกำหนดผู้รับผิดชอบ วิธีทดสอบ และการส่งมอบ ผลงานแต่ละโครงการมีขอบเขตแตกต่างกัน โปรดดูรายละเอียดในหน้าผลงาน
08 . further
ready to scope the architecture?
the paid technical discovery hands you the architecture spec and the eval plan before a line of the agent loop gets written. fixed fee, credited to the build, yours to keep either way. same studio that took propick property's google clicks from 39 to 79 and impressions from 1,163 to 3,955 in 28 days. brussels since 2021, sukhumvit since 2024, line official account open for thai-first contact.