NeuroSmith Technologies · AI engineering studio
AI that’s tough, not brittle.
A remote studio that designs, builds and hands over AI systems: LLM agents, assistants that answer from your own documents, computer vision and machine learning. Every milestone ships with its source code, a demo video and an install guide.
Plain facts
- ActiveSince January 2024
- TeamA team of four
- WhereRemote from Pakistan, clients worldwide
- WorkSeven systems on record, four of them under NDA
01 · Work · mill certificates
Systems we’ve built.
Some of our work is public. Much is under NDA. NDA certificates keep the engineering and drop the names.
| Certificate | What it is | Input stock: inputs and data | What we built process | Status |
|---|---|---|---|---|
NS-C04Under NDA
|
Claims-negotiation agentsAgents that read an insurer’s estimate, find the missing line items, write the argument and keep negotiating. | The insurer’s estimate, as a PDF | PDF extraction then Scope ontology flags omitted items and variances then LLM reasoning with safety post-processing then Variance report then Email worker that sends, replies and follows up |
Delivered · Sep 2026 |
|
||||
NS-C05Under NDA
|
Inspection QA from photosComputer-vision quality checks of field inspections, from photos taken on site. | Guided multi-view photos at each installation stage, captured inside a geofence | Multimodal model returns a verdict with a confidence and a summary then Branded PDF report Role-based, multi-tenant |
Delivered · Sep 2026 |
|
||||
NS-C06Under NDA
|
Enterprise RAGAsk the company’s documents a question; analytical answers come back with a chart they drew themselves. | Company documents in PDF, DOCX, XLSX, CSV and JSON | Ingest to S3 then Embeddings in a per-tenant vector index then LLM answer, plus an auto-generated chart titled in the asker’s language Role-based data rooms, shared chats |
Phases 1–5 delivered |
|
||||
NS-C03Public
|
WhyFashionA swipe-to-shop fashion app with virtual try-on. | Shoppers’ swipes, and the photos they search with | Swipe-to-learn recommendations, Search by image, AI virtual try-on In one app |
Live on App Store & Play Store · Jan 2026 |
|
||||
NS-C01Public
|
TuneTutorAn AI music-learning app: photograph your lesson notes, get back a song about them. | Photos of lesson notes | OCR then Level-aware summary then LLM-written lyrics then Generated song with karaoke lyrics In a Flutter app |
Delivered |
 
|
||||
NS-C02Public
|
Virtual MichaelA real-time talking-avatar travel concierge you can speak to. | A hand-built theme-park knowledge base of data sheets and reference files | LiveKit voice agent (speech-to-text, voice activity and turn detection, text-to-speech) then LlamaIndex RAG then Talking avatar In a Flutter app |
In progress · substantially built |
|
||||
NS-C07Under NDA
|
Agentic content newsroomA newsroom of agents: crews that plan, write, edit, translate and publish. | — | Planner then Writer then Editor then Translator crews Plus a title-optimisation crew; each agent on the model you choose; auto-publishing to WordPress |
Not stated |
|
||||
Status: In progress Delivered Live · The strip under each coupon repeats it: two stops in progress, four delivered, five live. · Metrics appear only with a client’s permission.
Want a certificate of your own?
Start a project02 · Services · the index
What we build.
Eight things we’re hired for, in alphabetical order, each linked to the work behind it.
-
AI Software-as-a-Service
AI inside a product you sell: multi-tenant, role-based, and built for paying customers rather than a demo.
Typical builds: a multi-tenant claims-negotiation platform; photo-based inspection QA with role-based access; document search with a separate index per tenant.
Certificates Claims-negotiation agentsNS-C04 Inspection QA from photosNS-C05 Enterprise RAGNS-C06
-
Chatbots
Assistants that answer from your own documents and say so when the documents don’t cover the question.
Typical builds: an assistant over company documents whose analytical answers arrive with a chart; a voice concierge that answers from a hand-built knowledge base (in progress).
Certificates Enterprise RAGNS-C06 Virtual MichaelNS-C02
-
Computer Vision
Systems that judge what is in a photo: inspection checks, visual search, virtual try-on.
Typical builds: inspection QA that turns guided photos into a verdict with a confidence and a report; search by image and virtual try-on in a shopping app.
Certificates Inspection QA from photosNS-C05 WhyFashionNS-C03
-
Deep Learning
Neural networks designed and trained for the problem in front of you, when an off-the-shelf model is not enough.
Specimen: the network at the top of this page, which learns our wordmark on your device.
Certificates No certificate listed yet See the bench
-
Generative AI
Text, image, audio and video generation built into a product, with the plumbing that makes it repeatable.
Typical builds: photographed lesson notes turned into a song with karaoke lyrics; virtual try-on before you buy.
Certificates TuneTutorNS-C01 WhyFashionNS-C03 Agentic content newsroomNS-C07
-
Large Language Models
LLM systems that read documents, reason over them and write, grounded in data you control.
Typical builds: agents that read an insurer’s estimate, find the missing line items and argue for them; retrieval over company documents in five file formats.
Certificates Claims-negotiation agentsNS-C04 Enterprise RAGNS-C06 Agentic content newsroomNS-C07 Specimen: ask about us
-
Machine Learning
Models that learn from your data to predict, rank or recommend, judged on data they have not seen.
Typical builds: swipe-to-learn recommendations in a fashion app that is live on the App Store and Play Store.
Certificates WhyFashionNS-C03
-
Virtual Agents
Agents that carry a task through many steps, or a conversation, by voice, email or chat.
Typical builds: a negotiation agent that sends, replies and follows up by email on its own; a talking-avatar travel concierge (in progress).
Certificates Claims-negotiation agentsNS-C04 Virtual MichaelNS-C02 Agentic content newsroomNS-C07
Specimen under Chatbots and Large Language Models
Ask about us.
With JavaScript on, you can type a question about NeuroSmith here. A BM25 search over this site’s own text returns the three passages a grounded assistant would cite, each with its score. When nothing scores high enough it answers: “I don’t know. Nothing on this site covers that.”
Have a system in mind?
Start a project03 · How we work
How a project runs.
Every project runs the same process and produces the same five things, listed below. If the details are sensitive, we sign an NDA before you share them. The illustrative brief beside the stages shows how a wish can become something you can check.
- 01Implementation plan
- 02Architecture document data flow, schema, API map
- 03Milestones each with its source code, a demo video and an install guide
- 04Revision rounds
- 05Hand-over documents
-
Discovery Assay
- What we do
- If the details are sensitive, sign an NDA before you share them. Read the brief and whatever data exists, then write the implementation plan.
- What you get
- 01Implementation plan
- What you decide
- Scope, milestones, and what “done” means.
Illustrative brief → tests
any question→ T1 ·about our company→ T2 ·instantly→ T4 Draft -
Prototype Forge
- What we do
- Design the system on paper: how data flows through it, the schema it lands in, and every API it calls or exposes. Then build towards the first milestone.
- What you get
- 02Architecture document: data flow, schema, API map
- What you decide
- Whether the architecture fits your systems before the build goes further.
Illustrative brief → tests
never wrong→ T3 Draft · T1, T2, T4 Measured -
Harden Temper
- What we do
- Deliver milestone by milestone, each with its source code, a demo video and an install guide. Fold your feedback in through revision rounds.
- What you get
- 03Milestones, each with source, a demo video and an install guide
04Revision rounds - What you decide
- What changes in each round.
Illustrative tests T1, T2, T4 Passing · T3 Measured
-
Hand over Hallmark
- What we do
- Hand the system over with its documents, so your team holds everything the project produced.
- What you get
- 05Hand-over documents, alongside the plan, the architecture document and every milestone’s source, demo video and install guide
- What you decide
- Sign-off.
Illustrative tests T1–T4 Passing
04 · Method · how we judge a model
We report the second number.
A model has two scores. The first is measured on the data it learned from, and it flatters. The second is measured on data it has never seen, and it tells you how the system will behave in your business. That is the number we work to.
Specimen · OverfittingComputed in your browser
This explorable fits a curve to noisy points with ridge regression on random Fourier features, solved exactly. A slider sets the model’s capacity. As capacity grows, the error on the training points keeps falling; the error on held-out validation points falls, then climbs again as the model starts memorising noise (and, past one feature per training point, it can fall again).
It needs JavaScript; nothing here is a recording.
- We judge trained models on data they never saw.When we train or tune a model, validation data is set aside first and stays aside.
- We say so when AI is the wrong tool.Sometimes a rule, a query or a form does the job better. We’ll tell you.
- We don’t publish numbers we can’t show.That is why this site has no results percentages: client metrics stay with clients.
For ML-literate readers: past the dashed line, where there is one feature per training point, the validation error peaks and then falls again. That is double descent. Mild ridge regularisation keeps the peak finite but does not remove it; the ring marks the lowest validation error on the data drawn.
05 · Team · maker’s marks
The four people behind the work.
Each NeuroSmith has a touchmark of their own.
Touchmarks · one per NeuroSmith
-
Safiullah Rahu
Chief Operations Officer
-
Munsif Raza
Lead LLM Engineer
-
Muhammad Talha
Full-Stack Developer
-
M Samama
AI Engineer
A small studio of named people: the four above are the people you work with.
06 · House rules
Standards and questions.
Standards
- No invented numbers.Every number on this site is a fact about us or our work, or is computed by the code on this page: live in your browser, or at build and labelled as recorded.
- Client names appear only with permission.That is why four of the seven certificates above are shown without names or screenshots.
- Live demos compute live; recorded runs are labelled.The network at the top trains on your device. When it can’t, you see a recorded run, and it says so.
- An NDA before details.If the details are sensitive, we sign an NDA before you share them.
- Every milestone ships whole.Each one comes with its source code, a demo video and an install guide.
Corrections: none yet. Last reviewed .
Questions
How does a project start?
You send a brief; the form below composes one for you. If the details are sensitive, we sign an NDA before you share them. Then we write the implementation plan, followed by the architecture document: data flow, schema and the API map.
Which time zone do you work in?
Pakistan Standard Time, UTC+5.
Studio time: Pakistan Standard Time
Will you sign an NDA?
Yes, and before any details are shared. Much of our work is under NDA, which is why several certificates on this page are shown without names or screenshots.
What do we get at hand-over?
Everything the project produced: the implementation plan; the architecture document (data flow, schema, API map); every milestone with its source code, a demo video and an install guide; and the hand-over documents.
Question not here? Ask it directly.
Email your question07 · Contact
Start a project.
Tell us what the system should decide or do. Write to us directly, or let the form below compose the email for you.
Hallmark of this visit
- Fig. 1 · The bench hasn’t trained on this visit.
- Fig. 2 · Not run on this visit.
- Fig. 3 · No questions asked yet.
Computed on this device. This page sends nothing you type; the brief form only opens your own email app.