layer 02 / mesh
Independent technical due diligence  /  Abu Dhabi

The pitch deck says unicorn. The code says weekend project.

I open the repository behind the pitch deck and tell you, in writing, whether there is a company underneath it. Six risk vectors, one score, five working days.

Book a 15 minute call Read a real audit report
Move the lens across the image. Look closely at the head.Look closely at the head
001  /  The premise

You can read the founder. I read the code.

Large funds keep a technologist on the bench for exactly one reason. Somebody has to open the repository, not to be impressed by it, but to establish whether the thing described in the deck exists in the form described.

You are making the same decisions without that person in the room. Not because you cannot judge a business, a founder or a market. You judge those better than most funds do.

It is that AI is the first category where the entire claim can live in a layer nobody at the table can read, and where the gap between a real system and a convincing one is three weeks of work.

That layer is the only thing I look at. I have no view on your thesis and no opinion on the founder. I tell you what is built.

002  /  The method

Six places the truth is kept.

Every engagement runs the same six vectors in the same order. Select one to see what the evidence actually looks like.

A repository keeps an honest diary. Commit cadence, authorship spread and the ratio of new code to rewritten code tell me whether a company was built over two years or assembled over one quarter to be shown to you.

COMMITS / MONTH MONTH 01MONTH 24 PRE-SEED ANNOUNCED 614 commits in 11 weeks 18 months, 1 author, 94 commits
Fig 01  /  Deal LQ-0231  /  git log --numstat, 24 month window
F-01 HIGHEighty one percent of the codebase was written in the eleven weeks before the raise, by a single author, with no review history. This is a demo built for a room, not a product built for users.

The demo serves twelve people beautifully. I model the point at which it stops, which component gives way first, and what it costs to fix. Most architectures do not fail gradually. They fail at a specific number.

P99 LATENCY 1001K10K100K CONCURRENT CLAIMED: FLAT TO 100K KNEE AT 8,400 p99 1.2s to 14.6s single write primary, no read replicas
Fig 02  /  Modelled from schema, queue config and connection limits
F-04 HIGHEvery inference request writes synchronously to one primary database with a connection ceiling of 200. The architecture cannot reach the traction the model assumes without a rebuild of the persistence layer.

One dependency with the wrong licence can place the intellectual property you are buying outside the company that is selling it to you. I walk the full dependency graph, including the parts the founders did not know they had.

app/ fastapi ml-core vectordb ingest pydantic MIT seg-attn-v2 GPL-3.0 torch BSD-3 faiss MIT httpx BSD-3 247 PACKAGES WALKED  /  4 DEPTH LEVELS  /  1 COPYLEFT CONFLICT
Fig 03  /  Transitive dependency graph, production lockfile
F-07 CRITICALA GPL-3.0 segmentation module sits four levels deep inside the inference path. Distributed as it is today, the copyleft terms reach the code the company calls proprietary. This is a legal question, not an engineering one, and it changes what you are buying.

AI margins invert quietly. A company can grow revenue and lose more money on every new customer, and the deck will not show it because the deck prices tokens at last year’s rate and this year’s volume. Move the inputs and watch where their model actually sits.

Gross margin per seat
-98%
Monthly inference bill
USD 459K
Negative unit economics at stated price
Fig 04  /  Live model, defaults are deal LQ-0231 as billed  ·  drag tokens to 9.5M to reach their heaviest decile
F-11 HIGHAt blended usage the margin is eighty percent, which is survivable. The heaviest decile consumes seven times the average and crosses into loss at 7.1 million tokens per seat. That is the cohort the sales team is actively targeting, so growth moves this the wrong way. The deck also models tokens at USD 1.80, a rate the company has never actually paid.

Where the training data came from, whether they had the right to use it, and what leaves the building. Two of the last nine companies I looked at were training on customer data their own contracts prohibited them from touching.

WEB CRAWL14.2M docs CLEANINGdedupe, pii scrub TRAIN SETv7.final.2 MODEL CLIENT UPLOADS3 enterprise accounts NO CONSENT PATH MSA clause 8.3 prohibits RETENTION: INDEFINITE  /  DELETION ENDPOINT: NOT IMPLEMENTED
Fig 05  /  Data lineage reconstructed from pipeline code and storage buckets
F-13 CRITICALDocuments uploaded by three named enterprise customers are flowing into the training set. The company’s own master agreement forbids it. This is a contractual and regulatory exposure that sits on the balance sheet you are buying into.

Fine tuned, or a system prompt in a trench coat. There is a test for this and it is not the demo. I compare what the deck claims is proprietary against what the repository can actually be shown to contain.

CLAIMED IN THE DECK FOUND IN THE REPOSITORY Proprietary domain model Fine tuning pipeline Proprietary eval harness Inference orchestration Not present Not present 12 assertions, last run 7 months ago prompt_v9.txt  +  1 API call SWITCHING COST TO A COMPETING MODEL: ESTIMATED 3 DAYS
Fig 06  /  Claim mapping against repository contents and training artefacts
F-16 HIGHNo fine tuning has ever been run. The product is a 4.2 kilobyte system prompt in front of a general model, wrapped in a good interface. The interface is real work and has value. It is not a technical moat, and the deck prices it as one.
003  /  The instrument

Every audit ends in one number.

The Luminaq Index scores the six vectors from zero to one hundred and resolves to one of three verdicts. The same instrument runs on every deal, which is the only reason deals can be compared to each other.

0/ 100
Vector breakdown
004  /  Evidence

This is the document you are paying for.

A real audit, redacted to protect the investor and the company. Nothing here is a template. Read it before you decide whether I am worth a call.

Luminaq  /  Technical Due DiligenceConfidential
Technical Due Diligence
Report

Prepared for , at the request of , in connection with a proposed seed investment in .

Verdict
Withdraw

Luminaq Index 31 / 100. Two critical findings materially change what is being sold. Detail at sections 4 and 6.

EngagementLQ-0231
Instructed11 FEB 2026
Delivered17 FEB 2026
Repository accessREAD ONLY, 4 DAYS
AuthorS. R. SINGH
Luminaq FZE LLC  /  Ajman, UAEPage 1 of 14
2. Index SummaryLQ-0231  /  Confidential
Index summary

Each vector is scored independently against the same rubric. No vector is weighted higher than another. A single critical finding caps the overall index at 40 regardless of performance elsewhere.

VectorScoreFlags
1. Codebase integrity221 high
2. Architecture scalability381 high
3. Licence and IP91 crit
4. Inference economics271 high
5. Data lineage141 crit
6. Model moat311 high

Composite: 31 / 100. The company presents as a research led AI business. The evidence supports an interface business with a competent front end team and no defensible technical position.

Luminaq FZE LLCPage 3 of 14
4. Licence and IP ExposureLQ-0231  /  Confidential
Finding F-07
Critical

Observation. The production lockfile resolves version 2.1.4 at dependency depth four, reached through the segmentation path in . The package is licensed GPL-3.0.

Why it matters. The module is linked into the inference binary the company distributes to on premise customers. On the plain reading of the licence, the copyleft obligation extends to the surrounding work. The company describes that surrounding work to investors as proprietary.

What the founders said. When asked, the CTO was not aware the package was present. It arrived transitively through a vision utility added in .

Remediation. Replace with an Apache-2.0 equivalent, or purchase a commercial licence from the author. Engineering estimate two to four weeks. The prior distribution history is a legal question and is outside my scope.

Effect on index. Vector 3 scored 9. Composite capped at 40.

Luminaq FZE LLCPage 7 of 14
5. Inference EconomicsLQ-0231  /  Confidential
Unit economics

The deck models blended inference at USD 1.80 per million tokens. Billing exports for the trailing three months show an actual blended rate of USD 4.10, driven by retry volume and an uncached system prompt resent on every call.

Per seat, monthlyDeckObserved
Tokens consumed0.4M1.4M
Inference cost$0.72$5.74
Price charged$29.00$29.00
Gross margin97%80%
Heavy decile marginnot modelled-98%

The blended figure is survivable. The distribution is not. The heaviest ten percent of users consume 61 percent of tokens and are the cohort the sales team is actively targeting.

Note. Caching the system prompt is a two day fix and recovers roughly 30 percent. I have flagged it to the founders regardless of the outcome of this engagement.

Luminaq FZE LLCPage 9 of 14
8. Founder Question ListLQ-0231  /  Confidential
Questions to ask

Put these to the founders directly. The wording matters. Each one is answerable in a sentence by a team that has done the work, and produces a long answer from a team that has not.

  1. Show me the training run. Which week, which hardware, what did it cost.
  2. If your model provider doubled prices on Monday, what is your plan on Tuesday.
  3. Who wrote the code between month one and month eighteen.
  4. Which customer data is in the training set, and where is the consent recorded.
  5. What is your gross margin on the heaviest ten percent of your users.
  6. Name the component you would have to rebuild to reach fifty thousand users.
  7. What is in the system prompt, and how long is it.

You are not testing their intelligence. You are testing whether the artefacts exist.

Luminaq FZE LLCPage 13 of 14
01 / 05
005  /  Free diagnostic

Wrapper or moat. Eight questions.

Answer for a deal you are actually looking at. Ninety seconds, no email, no download. It will not replace an audit and it will tell you whether you need one.

Question 01 of 08

Provisional read
0

Ask them these next
    Have me look at it properly
    006  /  The auditor

    You get me. That is the whole product.

    Satish Rohit Singh, technical auditor, Luminaq
    S. R. Singh  /  Luminaq  /  Abu Dhabi

    Founders can charm a generalist. They cannot hide a commit history from someone who has written one.

    I am Satish Rohit Singh. I have spent fifteen years building the systems I am now asked to examine, which is the only qualification that matters here. I was the first hire on a product team and co built that product from an empty repository. I have stood up data science functions inside companies that had none. I have shipped generative AI into production, with the cost reports and the incident reviews to go with it.

    That history is why this works. I know what a real training run costs, what a rushed architecture looks like six months in, and which corners get cut when a raise is three weeks out. I have cut some of them myself.

    I do not take equity, carry, or a finder’s fee, in your fund or in anything I audit. I have no relationship to protect on either side of the table. The only thing I am selling you is an accurate answer, including the ones that cost me repeat work.

    Background
    Fifteen years enterprise engineering and data architecture. Senior engineer through senior data scientist and technical product owner.
    AI specifically
    Four years in production generative AI. Retrieval systems, agentic architecture, evaluation, inference cost work.
    Founding side
    First hire and co builder on a venture backed product. Established data science functions at two companies.
    Academic
    MSc Data Science. Published peer reviewed research on explainable AI. Adjunct faculty teaching postgraduate AI.
    Independence
    No equity, no carry, no introduction fees. Registered in the UAE as Luminaq FZE LLC.
    007  /  Record

    The point is not the deals I stopped.

    Anyone can tell you no. What you are buying is the ability to say yes quickly, for a stated reason, and sleep afterwards.

    0
    Audits completed
    0
    Deals cleared with confidence
    0
    Median time to written verdict
    0
    Capital protected on stopped deals
    He found three things in four days that we had missed in six weeks. Two were fixable. One was not.
    Angel investor, Abu Dhabi
    Seed stage, MENA  /  deal proceeded after remediation
    Full attribution available under NDA on request
    008  /  Engagement

    Priced against the cheque.

    Diligence should cost a fraction of what it protects. The audit scales with the size of the decision behind it, so a small deal never subsidises a large one. You get the number before the work starts, never after.

    01  /  Signal
    AED 1,900
    48 hours  /  flat

    A first pass before you commit to full diligence. Enough to know whether the deal deserves the next four days of anyone’s time, including yours.

    • Repository and architecture skim
    • Wrapper or moat determination
    • Two page memo with a go or stop
    • Credited in full against a later audit
    Start a signal
    02  /  Full audit
    From AED 4,500
    5 working days  /  scaled to the cheque

    The complete instrument. Six vectors, scored, every finding evidenced and every claim in the deck tested against what the repository actually contains.

    • All six vectors, Luminaq Index score
    • Fourteen page written report
    • One page verdict for co investors
    • Sixty minute debrief call
    • Founder question list
    • One follow up round after remediation
    Book a full audit
    03  /  Standing
    AED 14,000/ month
    Priority queue

    For investors seeing deals continuously. You stop deciding whether a deal is worth diligence, because diligence is already retained.

    • One full audit or three signals monthly
    • 24 hour start on anything urgent
    • Unlimited short questions between deals
    • Quarterly portfolio technical review
    • Three month minimum, then monthly
    Discuss a retainer
    How I work
    Independence
    No equity, no carry, no introduction or success fees, in any company I audit or any fund I work for. I will decline an engagement rather than hold a position in it.
    Scope
    I establish what is technically true. I do not value the company, opine on the founders, or tell you whether to invest. That decision stays yours, which is where it belongs.
    No material finding, no feeNo equity, no carry Mutual NDA before accessRead only, time boxed access Clones destroyed at deliveryLuminaq FZE LLC  /  UAE
    Last thing

    Bring me the deal you cannot read.

    Fifteen minutes, no obligation, and I will tell you on the call whether it is worth auditing at all. Sometimes the answer is that it is not, and that is free.

    Book a 15 minute call Run the free diagnostic first