AI engineering · Faizan Hasnaat
I build AI that survives production.
For teams with real users, real payments, and real data to lose. Every output traces to its evidence, a human approves anything irreversible, and the pipeline heals itself when it breaks.
The sentence I am here for
The first one works.The second one needs building.There is one of me.
If that is your week, the rest of this page is for you. The first one works. The second one needs building. There is one of me.
01Shipped
Document extraction · in production
Receipts in.Books out.Bookkeeping that shows its work.
AI-powered bookkeeping for small businesses. Receipts and bank statements in — CSV, PDF, photographs — reconciled, tax-ready transactions out.
Every match carries a per-field confidence score you can audit, not a single blended number. Has paying customers.
Sources
- ⊢receipt_0912.jpg → OCR → 3 fields
- ⊢statement-aug.csv → 212 rows
- ⊢invoice_0417.pdf → text layer
92%
auto-matched · 8% to review
Tallyfold
illustrative
- STAPLES #1142$184.2098
- TELCO MONTHLY$62.1196
- SHELL 8807$39.9891
- AMZN MKTP$127.4574
⊢each match names the line it came from
Saidwhen
Tracks creators who make public stock calls and scores what those calls returned. Video in, transcript and on-screen text reconciled, structured claims out.
Every number traces to a named person, a named date, and a timestamp inside a video.
ColdDeck
Outbound that reads like you actually looked. A polite crawler finds a concrete problem on a prospect's site and drafts an email about that one thing.
Scoring is a deterministic rule pack you can replay — explicitly not a model. A human approves every send.
Read the case study →02What I do
Build it, then keep it working.
Three kinds of problem, and a page for each.
Build it
Something that does not exist yet, or exists as a prototype nobody trusts. Priced from a published floor rather than discovered on a call.
Keep it running
AI that already works, in a product that already has users, and nobody with the time to look after it.
Find out what is wrong
A feature that is live and getting things wrong, and nobody who can say why. Most engagements start here.
03Who you would work with
Faizan Hasnaat
The engineer who wrote it is the one on the call.
Engineer of Code and Proof. Five years shipping production software; the last two building AI systems where being wrong is expensive — extraction, retrieval and agent runtimes for teams who need to know why an answer came out the way it did.
04Questions I get asked
- What does a fractional AI engineer cost?
- Fractional AI engineering from Code and Proof starts at $2,500/mo. That buys ownership of the AI surface of your product — extraction, retrieval, evals, and the pipeline that runs them — at roughly two days a week, async, month to month with no notice period.
- What is not included in the retainer?
- Frontend work outside the AI features, model training or fine-tuning, and anything requiring a US entity or a security review. Building a new product from scratch is also outside the retainer, but only because it is a separate offer: that is AI application development, from $15,000. The exclusions are written down rather than negotiated later, because a retainer without an edge becomes a full-time job at part-time money.
- What does Code and Proof do?
- Code and Proof is a fractional AI engineering practice: one senior engineer who owns the AI surface of a live product month to month — extraction, retrieval, and agent runtimes — for teams with real users, real payments, and real data to lose. Every output traces to its evidence, a human approves anything irreversible, and the pipeline heals itself when it breaks.
- What does an AI reliability audit cost?
- An AI reliability audit from Code and Proof is $2,500 for five business days, at fixed scope and fixed price. It returns a written list of what will break, why, and what fixing it costs.
- What does the audit actually review?
- Six things: every place the system can be confidently wrong, what you measure today against what would catch a regression, whether a given output can be explained from your logs alone, everything the system can do that cannot be undone, what the pipeline costs at ten times current throughput, and what happens when it breaks at 3am.
- How do you stop an AI system from being confidently wrong?
- The model reads messy input, but it does not get to be the judge. Scoring runs on a rule pack that can be replayed and retuned, and every record carries the page, timestamp, or frame it came from — so any number the system produces can be pointed at rather than trusted.
- What stops an AI agent from doing something irreversible?
- A human gate. Nothing that cannot be undone — send, charge, delete, publish, file — happens unsupervised. The system drafts, and a person approves before it acts.
- What if my AI feature is already live and wrong?
- That is AI project rescue, and it starts at $15,000, scoped from what the audit finds. The $2,500 audit comes first, so the scope of the repair is based on evidence rather than an estimate. If the feature works and the problem is that nobody has time to keep it working, the retainer is the cheaper answer.
- Who does the work?
- Faizan Hasnaat, the engineer who writes the code, is the person on the call. Five years shipping production software, the last two building AI systems where being wrong is expensive.
- What does it cost to build an AI application?
- AI application development from Code and Proof starts at $15,000. That covers an AI product built end to end — the pipeline, the evals, the review surface a person actually uses, and the operational spine that keeps it running. The figure matches the published cost guide on this site rather than undercutting it: a thin AI feature runs $8k-$20k, a real pipeline $25k-$60k.
- What does it cost to build an AI agent?
- AI agent development from Code and Proof starts at $25,000, which is the agentic band in this site's own published cost guide. An agent is more expensive than a pipeline because the expensive part is not the model — it is the approval gate, the idempotency keys, the cost ceilings and the transcript, all of which exist to make being wrong survivable.
- What does an AI agent need before it can act in production?
- A human gate, enforced in the runtime rather than requested in a prompt. Sending, charging, deleting, publishing and filing all wait for a person to approve them. Every side effect carries an idempotency key so a retry is a no-op rather than a second invoice, and a per-run cost ceiling stops a loop becoming a billing incident.
- How long does an AI MVP take to build?
- An AI MVP from Code and Proof starts at $15,000 and is scoped to one workflow that works end to end rather than five that demo. The evals are written before the feature, so the second version can be measured against the first instead of argued about. If the scope cannot be cut to one workflow, it is not an MVP and it is priced as an application build.
- How much does an AI developer cost?
- Code and Proof publishes its floors: $2,500 for a five-day fixed-scope reliability audit, $2,500/mo for a fractional retainer at roughly two days a week, from $15,000 to build an AI application or MVP, from $25,000 for an agent that acts in production, and from $15,000 to rescue a feature that is already live and wrong. Large builds are scoped rather than listed, but nothing starts below these numbers.
- Do you work fixed-price or hourly?
- Fixed scope and fixed price for the audit, a monthly fee for the retainer, and a scoped fixed price for builds. Code and Proof does not bill hourly, because an hourly rate is a number to negotiate down and because being fast should not reduce the invoice.
- How do you add AI to software that already has users?
- By splitting the work in two. Reading, drafting, summarising and routing run unsupervised, because the worst case is a wasted minute. Anything irreversible waits behind a human gate. Side effects get idempotency keys so a redeploy or a duplicate webhook cannot charge twice, retries are scoped to the unit of work rather than the whole run, and cost ceilings are enforced in the runtime.
- Why does our RAG system return wrong answers?
- Almost always retrieval, not the model. If the right chunk never reaches the context window, no prompt will recover it. The fix starts with measuring recall separately from answer quality against a golden set, because a single blended score cannot tell you which half is broken — and then holding it there with a regression gate in CI.
- Will you approach our client directly?
- No, and it is contractual rather than a promise. On white-label work Code and Proof builds the AI feature under the agency's brand, does not sell the surrounding work into the account, and does not approach the end client.
05Start
Five days from nowyou’ll know
what breaks, why, and what fixing it costs.
$2,500 · fixed scope · five business days
Builds start at $15,000 and are scoped from what the audit found, so neither of us is guessing. If nothing is wrong, that is the finding. Or email codeandproof@gmail.com.