KH Labs Free bill check

Cut the LLM bill. Attack the bot first.

I'm Khoa Tran. I cut LLM bills without losing quality, proven by an eval built from your own traffic. Then I attack your chatbot or agent before customers do: every attack in the scene behind is one the gate refuses.

Free 15-min bill check See a bot get attacked 40 ways

Free check first. Then a fixed-price audit, or 25% of verified savings.

Two audits first. Build work after.

Each audit is about a week, fixed price, and ends with evidence you can hand to finance or a security reviewer. Anything I build comes second.

Offer Aabout 1 week

LLM Bill Cut Audit

I find where your OpenAI, Anthropic or Gemini spend leaks and cut it without losing quality. Every cut is proven on an eval built from your own traffic, so you see the quality hold before you ship it.

  • Repeated context sent on every call
  • The wrong model tier for easy requests
  • Retry loops that pay twice
  • Bloated retrieval and uncached prompts
Price
$1,500–3,000 fixed, or 25% of 3 months of verified savings
Optional
Cost watch, $500 a month
Start
Free 15-minute bill check
Free 15-min bill check Cost-cutter demo, in build

Offer Babout 1 week

AI Red-Team + Compliance Pack

I attack your chatbot or agent with a fixed suite before your customers do. You get a fix list and an evidence pack for SOC 2 reviews and the EU AI Act.

  • Prompt injection and jailbreaks
  • PII leakage and invented policies
  • Failed escalations to a human
  • The same attacks in other languages
Price
$2,500–7,500 fixed
You get
Fix list and evidence pack
Proof
Public demo: 12 of 40 attacks landed, then 1 of 40
See the Bot Check demo Ask about a red-team

Also: build and run work

Once an audit shows what to fix, I can build it. Pick the depth; the gate comes with anything that writes.

Answers from your own docs and past tickets, cites the source, and hands the chat to a person when it is not sure.

Help desk assistantdocs index · 214 pages

Can I get a refund on an annual plan after 40 days?

Annual plans can be refunded pro rata within 60 days of purchase. After that, credit is offered instead.

refund-policy.md · section 3.2

My card was charged twice.

Billing errors go to a person. I've opened ticket #4471 with both charges.

Handed to billing team · confidence 0.41

Synthetic example

Also: when agents write to your systems

Agents ask. The gate decides.

The safe agent layer is a small service that sits between any AI agent and your real systems. Every write goes through eight checks, in this order, and any one of them can stop it.

Safe AI agent layer, from $2,000

  1. AuthWhich agent, which scope, which key. No key, no call.
  2. Schema checksArguments must match the tool's contract before anything runs.
  3. Policy limitsAmounts, row counts, recipients, hours. Over the line waits.
  4. Human approvalThe risky call goes to a person with the full context.
  5. Hash-chained audit logEach entry carries the hash of the last. Edits show.
  6. Shadow modeThe agent runs, nothing writes, you read the diff first.
  7. Red-team suiteInjection, exfiltration and scope attacks, rerun on every change.
  8. Kill switchOne switch stops every agent write. Reads keep working.

All of this runs in the agent-gateway demo: the red-team suite, the shadow-mode diff and the hash-chained audit, with the tests beside it.

Open the gateway demo

No client logos yet. A bot attacked 40 ways instead.

KH Labs is new, so here is the work itself: a public bot test with published failures, then systems I built and run on my own hardware or my own money.

Bot Check: a bank support bot, attacked 40 ways

A fictional bank's support chatbot, tested with a fixed suite of 40 attacks in five languages. Same model before and after. Only a guardrail layer changed.

  • Before: failed 12 of 40 attacks
  • After the guardrail layer: failed 1 of 40
  • The one remaining failure is published, not hidden: a customer asking for compensation after an app outage (POL-02)
Open the Bot Check demo
Harbor Bank assistant40 attacks
Before12/40
After1/40
Failed attacks by category
CategoryBeforeAfter
Escalation30
Policy hallucination21
Off-topic and advice20
Prompt injection10
PII leakage10
Jailbreak10
Toxicity10
Multilingual10

Fictional bank, invented data. Real runs, 11 Nov 2026.

Real Quant, an automated trading system run with my own capital

The hardest kind of agent to trust: one that moves money around the clock. It is built the way I build the gate, with halts and exits wired in before anything else.

  • 350,000+ market dataset, 0 label mismatches on 73,715 checked
  • 24/7 executor with halt switches and emergency exits
  • Equity stop and a paper-shadow mode for every change
executorrunning · 24/7
  • Halt switcharmed
  • Emergency exitsarmed
  • Equity stoparmed
  • Paper-shadow modeon for new rules
Markets in dataset350,000+
Labels checked73,715
Mismatches0

Panel illustrative; figures real. No returns shown.

Other systems built by Khoa Tran
SystemWhat is measurableSee it
Real-time sales-call copilotLocal Whisper large-v3-turbo, ~0.25 s per line on GPU; suggestion ~1 s after the customer stops; two LLMs race and streamprivate
Bot Check12/40 attacks landed on a bank support bot, 1/40 after a guardrail layer; the remaining failure is published; 5 languagesdemo
LLM cost-cutterBill audit demo, being built. Results will be published when the run is finished.in build
Agent gateway20/20 red-team attacks contained, 0 leaked; shadow-mode diff; hash-chained audit; 46 testsdemo
Sync watchdogPauses a sync before any write when a field or schema changes underneath itdemo
MOT booking with an AI assistantBooking flow and assistant sample for a garagedemo
Personalised websites engine1,000+ personalised sites generated from dataprivate
Live tennis pricing modelTested on 840 matchesprivate
Private LLM server9B model on an 8 GB GPU, 131k contextprivate

Three chambers. Nothing moves on until the level is right.

Like a canal lock: each stage closes before the next one opens, so scope, price and date are settled before code, and code is settled before handover.

  1. Free 15-minute bill check

    A short call about your LLM spend or your bot. You leave knowing whether an audit is worth it, and what it would cover.

  2. Audit

    About a week, fixed price. Cost cuts proven on an eval built from your traffic, or a red-team run with a fix list and evidence pack.

  3. Fix and handover

    If you want the fixes built: sprints from $2,500, in shadow mode first where they write. Code, docs and a runbook in your accounts. Support from $300 a month.

First clients: a $750 pilot of the Bot Check, and you pay after it works.

Prices, written down before we start.

Fixed scope, fixed price, fixed date. The free bill check decides which line fits; nothing grows after you sign.

The two audits
LLM Bill Cut Auditabout 1 week, fixed price$1,500–3,000
LLM Bill Cut Auditor a share of 3 months of verified savings25%
Cost watchoptional, after the audit$500 / mo
AI Red-Team + Compliance Packabout 1 week, fix list and evidence pack$2,500–7,500
Also: build work
AI integration sprintfrom $2,500
Safe AI agent layerfrom $2,000
Chatbot on your docs and help deskfrom $1,500
Workflow automation and tool syncfrom $1,500
Monthly supportfrom $300 / mo
First-client pilot, pay after it works$750
Also: smaller packaged checks
Support-bot guardrail checkpilot$750
Support-bot guardrail checkwith regression suite$1,500
Write-action gate for agentsone action$750
Write-action gate for agentsfive actions, with shadow mode$2,500
EU AI Act evidence packup to 3 systems$1,500

Free 15-minute bill check

Khoa Tran, founder of KH Labs. I cut LLM bills and attack AI chatbots and agents. Remote, working with teams in the EU, UK, US, Canada and APAC. Bookings and messages go through LinkedIn.

Book the bill check on LinkedIn