Systems architect & CTO · Software and AI
I analyse, design and build complex software — SaaS platforms, applications and AI systems — and take them from an idea to a product people rely on every day.
Bring me the goal. Analysis, design, build, launch and support are mine to carry. Most recent: elEMZA, an e-signature platform, from a one-page brief to a live product in 70 days.
| Drawn by | Arash Banaeian | ||
|---|---|---|---|
| Title | Drawing set of a systems architect | ||
| Latest proof | 5.9× elEMZA contracts, against the old platform’s best month | ||
| Date | 2026-09-25 | Scale | 1:1 |
| Location | Remote · UTC+3:30 | Sheet | 00 of 08 |
Sheet 01 · Who I am
The eye is our oldest symbol of awareness. This one has been taken apart, layer by layer, and each layer is something that shaped the way I see. Light leaves a distant star, crosses every layer, and arrives as understanding. Every problem I take on travels the same road.
| 1 |
Rigel, in Orion Cosmos It began with the night sky. A teenager who spent hours staring at Orion learned one thing before anything else: see the whole picture first, then the details. |
|---|---|
| 2 |
Cornea Science Light meets the cornea first. For me, that layer is science: the first filter every claim has to pass. I don’t believe what I haven’t measured. |
| 3 |
Iris and pupil Creativity No two irises are alike. The light is the same for everyone; what each eye makes of it is not. I rarely reach for the ready-made answer. I design my own. |
| 4 |
Lens Depth The lens grows in layers for a lifetime and never sheds one. Every project, and every failure, has added a layer to the way I see, and sharpened it. |
| 5 |
Retina Intuition The retina is a piece of brain that grew outward to meet the world. Intuition works the same way: the answer arrives before its reasons. Then I measure until the reasons catch up. |
| 6 |
Visual cortex Mind In the visual cortex, points of light become lines and shapes. A mind does the same with the world, turning scattered signals into patterns. That is where my work begins: with the pattern no one has named yet. |
| 7 |
The recognition pathway AI Artificial neural networks were modelled on this pathway, and their first layer learns, untaught, the very edges the brain detects. I build machines to understand what thinking is, and perhaps where a mind ends. |
| 8 |
The observer Self-awareness And at the end, someone who sees. Science still cannot say where that inner experience comes from, or whether the red I see is the red you see. It is the one part no one has drawn yet. |
Hover or tap a layer to open it.
Sheet 02 · Growth, in numbers
The old platform was signing about 340 contracts a month until it failed: in January 2026 it recorded four. I was asked to build a new one, properly, from nothing. Seventy days later it went live, and every month since has beaten the month before.
Measured on production, 30 Sep 2026
Rated capacity · one 4-core server
≈ 61,000 contracts a day, sustained
All 2,118 September contracts fit in 49 minutes of it.
Load lab, 28 Aug 2026 · 4 cores · 7.9 GB
Hover a building for its figure.
Sheet 03 · Case study
elEMZA is where contracts and forms get signed online, with full legal force. I built it from nothing, from a one-page brief. Each layer of this contract was its own problem; open one to see how it was solved.
1Your own document
People upload the contract or form they already have — a photo, a scan or a PDF. Nothing has to be retyped.
2The clauses
An AI legal assistant drafts the clauses, critiques its own draft and checks every legal reference before it reaches the page.
3Where to sign
The place to sign, and every field to fill in, is marked precisely on the document itself.
4Identity
Every signer is checked against the national registry, with a face match: 4,600 live checks so far.
5Digital signature
A legally valid signature from a licensed certificate authority, sealed inside the PDF itself.
6Timestamp
An independent authority, not our own server clock, certifies the moment of signing.
7Audit trail
Every action is logged with who, when and from where: a record that holds up in court.
8Connected systems
The customer’s own software hears about the signature the moment it happens, through an API, a PHP kit and a WordPress plugin.
Hover or tap a layer to light it on the document.
70 days, first commit to relaunch
First contract on the new platform; first signature 42 seconds later.
Sheet 04 · The team I built
elEMZA was built by an engineering team made entirely of AI agents. Each agent has its own role, memory, rules and duties; each group has a coordinator that drives its work. I design, oversee and tune the whole ensemble, the way a conductor leads an orchestra. Nothing they produce reaches the live system without passing a strict control system I built. Every unit on this drawing is one of those controls.
Click any unit, or a row of the list.
The work that passed through
| TK-01 | 5–15 a day | |
|---|---|---|
| F-02 | 23 files | |
| V-17 | 17 hooks | |
| PI-03 | every claim | |
| R-04 | every change | |
| CV-05 | 7 deploy assertions | |
| TK-06 | 100 entries | |
| TK-07 | 99.97% without a server error |
Sheet 05 · Learning from failure
The system learns the way a reinforcement learner does: every failure is measured, traced to its root and turned into a rule that makes a repeat structurally impossible. The result is an agent team that works more precisely each week than the week before. Four examples:
Lessons recorded100entries · counted 24 Sep 2026
Click a tremor to open its record.
The ledger now holds a hundred entries. The first is the day a one-line command erased sixteen files. I keep it at the top on purpose: a standing reminder of what confidence without measurement costs.
What happened A one-line transform through a shell hit an escaping quirk: the regex failed to compile, the replace returned null, and sixteen source files were written as zero bytes and committed.
Why it slipped through The syntax checker passed them; an empty file has no syntax errors.
The rule it learned The catalogue’s first entry. A zero-size scan is now part of every bulk edit.
What happened A component shipped with a public property and a public method of the same name. In the browser the click resolved to the property and did nothing.
Why it slipped through Every test was green, because tests call the method directly.
The rule it learned The fix was not the rename; it was a test that scans every component for the collision, with a planted one to prove it fires.
What happened A layout setting was added, a test asserted the setting, it was deployed — and nothing changed on screen. The framework reads it from a different class.
Why it slipped through The test asserted the input, not the output.
The rule it learned Assert the rendered output, never the input; before trusting a setting, find the line that reads it.
What happened Image generation had been failing for weeks on a valid key sent to the wrong host.
Why it slipped through The job wrote a receipt on success and nothing on failure, so “never attempted” and “attempted and failed” were the same empty query.
The rule it learned Every terminal path writes a receipt: success, refusal and failure.
Sheet 06 · How I work
In every new field I first learn how its people actually decide and behave, then build the system around that. The fields change; the skeleton doesn’t: take in, model, decide, act.
Pick one of the four systems to see it laid over the same skeleton.
1 Elicit · what the person gives
Intake, family interview, six psychometric instruments
The learner’s question, history and level
Conversation, voice, product photos, receipts
Form fields on scanned originals; identity and face match
2 Model · how the system holds it
Scored profile joined to qualitative case notes
Textbook rebuilt as structured knowledge; learner memory
Layered memory; intent, sentiment, personality, journey state
Contracts, parties, roles, organisations, audit trail
3 Decide · what it chooses
Career-fit matching; a three-advisor commission sets three paths
Hybrid retrieval (dense + BM25, RRF, reranking); adaptive path
Adaptive sales strategy under live A/B tests
Signing workflow with provider health gates
4 Act · what it does
Final consultation and report; 4,000+ people assessed
Grounded answer citing page and paragraph
Quote, confirmation, payment evidence, human hand-off
PAdES signature via a licensed CA; API and signed webhooks
Sheet 07 · Star chart
At sixteen I became the youngest paid astronomy instructor in Mashhad. I still approach every new field like a sky no one has charted yet.
β Ori · Rigel
Regulated e-signature, identity and PKI; sole technical owner, from brief to production.
The real stars of Orion, at their real positions and brightness; each carries one domain. Click a star.
Orion’s altitude over Mashhad, tonight
Magnitude
RA: hours · Dec: degrees · J2000
Sheet 08 · Last sheet
| arash.bne@gmail.com | |
| Telegram | @eArash |
| linkedin.com/in/arash-banaeian | |
| Location | Remote · UTC+3:30 · English and Persian |
| Documents | CV · PDF · Case study · PDF |
Approval block
Architect
Arash Banaeian
Client · you
Sign on the line with a mouse or a finger. Your drawing is never stored or sent; only the fact that you signed is counted.
| Rev | Description | Date | By |
|---|---|---|---|
| 0 | Drawn and checked | 2026-09-25 | A. B. |
| A | Awaiting your signature | — | — |
What you get: the system that makes AI output safe enough to sign a contract with, and the person who wrote it. It moves to your codebase in days, and someone who has never met me can audit every rule in it.