Production-grade AI, not a demo. Shipped in 14 days at a fixed price.

We agree on the scope and pass/fail checks before coding. Your code is in your GitHub from day one, and the build is done only when those checks pass in production.

Send the brief See what ships

Fixed: price and scope in writing
Day 1: code lands in your GitHub
Done: written checks pass in production

I'm Simon, founder of Keelfast. I own your build. A second engineer cross-reviews it, and QA checks the release.

Two systems. The important parts stay visible.

These are reconstructions of shipped workflows with synthetic data. The implementation claims come from source review, not from the artwork.

ALL FRAMES · ANONYMIZED RECONSTRUCTION · DEMO DATA

Enterprise RAG

Document search · OCR · metered AI

Years of contracts and scanned paper had to become searchable without turning the answer into a black box. The system keeps retrieval, source inspection and usage accounting visible to the operator.

Input: Direct text extraction or OCR
Retrieval: Qdrant + Meilisearch, RRF + reranker
Control: File inspection + idempotent token ledger

What shipped

Scans and text documents enter one searchable corpus
Vector and keyword result sets are fused, then reranked
Each AI run records tokens, price snapshots and its idempotency key

FastAPI · Qdrant · Meilisearch · RabbitMQ

Anonymized Enterprise RAG reconstruction showing separate vector and keyword result lists fused with RRF and reranked into a final evidence set — **FIG. 02** · ENTERPRISE RAG · HYBRID RETRIEVAL + RERANKING

Inspect ingestion and metered usage

Anonymized Enterprise RAG reconstruction showing a scanned PDF taking the OCR route before chunking and indexing — **FIG. 03** · INGESTION · OCR ROUTE

Anonymized Enterprise RAG reconstruction showing token counts, price snapshots and duplicate-charge protection in a ledger — **FIG. 04** · USAGE · IDEMPOTENT LEDGER

AI Visibility

Research SaaS · six source paths · resilient collection

A founder needed an inspectable answer to what AI and search products say about a brand. The product preserves the wording, route and limitation of each source instead of hiding everything behind one score.

Coverage: Six implemented answer and search paths
Failure: Per-source circuits, partial reports, ChatGPT fallback
Release: Configured CI gate + tested bounded restore drill

What shipped

Each observation keeps its source wording and collection route
Completed paths remain inspectable when another source stops
Tracked questions stay editable before a manual analysis launch

FastAPI · Next.js · OpenAI API · Docker

Anonymized AI Visibility reconstruction showing a concise finding, six source paths and a ChatGPT-specific fallback route — **FIG. 05** · AI VISIBILITY · FINDING + SOURCE ROUTES

Inspect source evidence and release controls

Anonymized AI Visibility reconstruction showing the exact wording, provenance and limitation of one source observation — **FIG. 06** · OBSERVATION · SOURCE WORDING

Anonymized AI Visibility reconstruction showing auth, billing, configured coverage, immutable release and tested restore controls — **FIG. 07** · RELEASE · IMPLEMENTATION EVIDENCE

Production AI MVP in 14 days

The hard part starts after the first model response: bad output, provider failures, changing API costs and real users. We build the controls around the AI at the same time as the product, then test both against one written definition of done.

AI that fails safely

Schema-validated outputs, fallbacks, retries, tracing and measured cost per request. The critical path is checked against written acceptance cases before launch.

A real product around it

Auth, payments, integrations, deployment, monitoring and tests. These are the parts that turn a model call into software people can use.

A handoff you can own

Code in your GitHub from day one, plus tests, a README, a runbook and a cost forecast based on measured requests. You keep the whole system.

Day by day, in writing

One call at the start. After that, a written update every workday and a short demo video every 2–3 days. Progress, decisions and blockers stay visible without another status call.

Day 0The only call of the project
30–45 minutes. The same day you get the fixed price, the spec, the acceptance checklist in writing, and your GitHub repo.
Days 1-7Foundation, then the AI core
Architecture, CI/CD, auth, data model and a staging deploy. Then the AI pipeline with structured outputs, fallbacks and cost tracking. The first written update lands on day 1; demo videos come from a live staging site.
Days 8-13The rest of the product, then hardening
The remaining features, then a security pass, monitoring and the agreed checks. On day 13 you review the product against the same checklist, and we fix what does not match.
Day 14Production launch
Live on your domain. You receive the docs, runbook and measured LLM cost forecast. Then the 30-day bug-fix window begins.

Already have a Lovable, Bolt, or v0 prototype? We can keep what works and harden the path to production: security, real auth, tests, deployment and monitoring. The same written scope and acceptance process still applies.

You read the project as it ships

A fixed-price project should still be easy to inspect. Every build uses the same three written artifacts. These are the actual formats with clearly illustrative content.

Daily updateWhat changed, what is next, and what needs a decision.

SimonExample · day 8 of 14 · 17:46On track

Done today

Implemented output validation and safe retry behavior
Added the operator review step before anything is sent
Covered the new path with automated tests

Connect the staging form to the client workspace and prepare tomorrow's acceptance run.

Decision needed

The CRM sandbox still returns placeholder contact IDs. I can proceed with fixtures today; real API access is needed before day 10.

Client 18:08
Fixtures are fine for now. I'll send sandbox access tomorrow morning.

Four lines, every workday, without you asking.

Acceptance checklistThe same pass/fail definition from day 0 to release.

Day-0 agreementexample run · 14:323 of 4 passed

Intake produces a structured briefPassrun_1842
Operator approves before sendPassvideo · 01:14
Duplicate webhook does not charge twicePasstest_retry_07
Client can export an audit recordReviewstaging link

Agreed on day 0. On day 13 you check the product against it.

LLM cost forecastMeasured cost per run, then your volume assumptions.

fx = MONTHLY_RUNS × ESTIMATED_COST_PER_RUN

Scenario	Volume input	Forecast state
Measured proof	sample run	baseline
Launch	expected volume	estimate
Guardrail	upper bound	alert

Excludes infrastructure and support; those are budgeted separately in the SOW.

Example format only. Your inputs are measured from the agreed proof before launch.

One scoping call at the start. After that, decisions stay in writing and the repository remains open to you.

If we miss the agreed 14-day deadline, we keep working for free until it's done. And done is defined in writing before the start: the build passes the agreed acceptance checks and deploys to your production environment.

acceptance checks agreed in writing
deadline in the contract
code in your GitHub from day one

One fixed build. One smaller first step.

You know the full price before we start. If the build still has a technical unknown, the sprint resolves that first and its $750 counts toward the MVP.

AI Feasibility Sprint

$750

~5 days · paid upfront

A clear answer before you spend real money.

Feasibility report: what to build, how, and what the risks are
Working proof of the riskiest AI part, delivered as code in your repo
LLM cost forecast for your real volumes
A fixed quote for the 14-day build; the price will not change later
The $750 counts toward the build price

Send the brief

Production AI MVP

from $3,999

14 days · fixed price · 50% upfront

A product that survives real users.

Up to 5 features, including the AI core with its production layer
Auth, payments, deploy, monitoring, tests
A daily written update, plus a demo video every 2-3 days
Code in your GitHub from day one
30 days of free bug fixes after launch
Covered in full by the Keelfast guarantee: deadline and production

We take one MVP build per month, so one project gets full focus.

Send the brief

What from $3,999 covers

Up to 5 features with one AI core and its production layer, plus auth, payments, deploy, monitoring, and tests.

When the quote grows

More features, extra integrations, a data migration, several user roles, or compliance work. Whatever the scope, you see the exact number before we start.

Care Plan from $500/mo: monitoring, cost alerts, bug fixes, small improvements.
Iteration Plan from $1,900/mo: we keep shipping your roadmap month by month.

Simon, founder of Keelfast, in a natural-light editorial portrait — **FIG. 06** · SIMON · FOUNDER, KEELFAST

I'm Simon, founder of Keelfast

7+years in software engineering
14client projects delivered before the studio

Keelfast is a studio of four: three engineers and QA. One senior engineer owns your build end to end, a second engineer cross-reviews it, and QA checks the release. You always talk to the person responsible for the code.

The fourteen client projects behind this page are mine, delivered before the studio existed. We now work async-first: one scoping call, then written updates and short demo videos. I sign off on every release.

The studio behind every delivery

ALIKHANENGINEERING
OLEGENGINEERING
ELIZAQA · DESIGN REVIEW

Fair questions

Will you really finish in 14 days?

The deadline is in the contract: if we miss it, we keep working for free until it's done. You see progress through a written update each workday, a demo video every 2–3 days, and the code in your GitHub from day one.

Who actually works on my project?

Keelfast is four people: three engineers and QA. One senior engineer owns your build end to end. A second engineer cross-reviews the critical changes, QA runs the release checklist, and I sign off on the release.

I don't want to pay upfront to someone I don't know.

Fair concern. The work is split into milestones (50% to start, 30% after the day-8 demo, 20% at launch), so you approve progress before each payment. And the code lives in your GitHub from day one: if we disappear tomorrow, you keep everything.

What fits in the 14-day build?

The base scope covers one AI core and up to five product features, including auth, payments, deployment, monitoring and tests. Extra integrations, several user roles, data migration or compliance work can increase the quote. If the riskiest technical part is still unknown, the $750 Feasibility Sprint resolves it first and counts toward the build price.

Is AI-generated code safe? Who will maintain it?

AI helps us implement faster; it does not decide what is safe to ship. Changes are reviewed, the critical paths are tested, and the test suite and CI live in your own repository so another developer can take over.

What happens after the 14 days?

You get 30 days of free bug fixes, plus docs and a runbook. After that you have three options: take over fully (the code and infrastructure are already yours), move to a monthly care plan for monitoring, fixes, and small improvements, or start the Iteration Plan, where we keep shipping your roadmap month by month.

Send the brief

Start on Upwork so identity, messages, milestones and payment protection stay in one place. A few lines are enough; nothing is charged before the scope and price are agreed in writing.

Four useful lines beat a long pitch.

Send the brief

Opens Simon's public Upwork profile. You can start in writing; no call is required.

Current state: Idea, spec, prototype, or live product
Needed outcome: The one workflow that must work
Deadline: A real date, or “no hard deadline”
Useful link: Prototype, repository, or short spec (optional)