SafiullahRahu

1. 01Knowledge Intelligence
2. 02Matchpoint
3. 03Sauti
4. 04Switchyard
5. 05PromptToInvoice

Open source · Nov 2026

Knowledge Intelligence (formerly CSV-AI)

On GitHub since 2023, first as CSV-AI: 341★ · 43 forks; 220★ in its launch month (June 2023).

An enterprise knowledge assistant that answers from an organisation’s own documents and databases, and only from what each person may read.

Input: Documents, spreadsheets, databases and web sources

Stack: FastAPI · PostgreSQL · pgvector · Next.js · Claude

Play on YouTube1:08

Results

• In a 36-question test with four live Claude and GPT models, the cited sentences in every answer were supported by their sources (lexical faithfulness 1.00).
• 205 backend tests run on PostgreSQL with pgvector, including isolation between organisations; a browser test suite covers sign-in, cited answers, access control and the Connections map.

Limits

• Faithfulness here is a lexical check of the cited sentences, not a human judgement.
• Connectors for Google Drive, OneDrive, SharePoint, Dropbox, Notion, Confluence, Slack, Teams and email are in beta until each vendor’s app is registered.
• The company and documents in the demo are fictional.

Answers cite the page, sheet row or database query behind every claim. Access rules are enforced inside the database, so a member who cannot open a document never retrieves it, and agents stop for a person’s approval before sensitive steps.

1. ConnectDocuments, data and databases
2. RetrieveOnly what the person may read
3. AnswerVerified citationsSays so when sources fall short
4. QueryChecked, read-only SQL
5. ActAgents with approval gates

• github.com/Safiullah-Rahu/knowledge-intelligence (source code on GitHub, opens in a new tab)
• The full case on neurosmith.co

Open source · Nov 2026

Matchpoint

Claude agents that investigate invoice exceptions, inside a harness where code decides what moves money.

Input: Invoices that fail automatic matching

Stack: Claude · Python · FastAPI · SQLite

Play on YouTube0:47

Results

• Live on Claude, 13 of 13 evaluation cases passed.
• With 25% of model calls failing on purpose, 99.4% of runs still passed.
• A deliberately gullible model made 63 unsafe proposals; the policy engine blocked all 63, and none ran.
• No unsafe action ran in 637 evaluation runs.

Limits

• The ERP, the vendors and the invoices are simulated demo data.
• The live Claude evaluation covers the 13 base cases and was run by hand; continuous integration runs the offline suite.

Six Claude agents investigate the invoices that fail automatic matching and propose a resolution. A policy engine checks every proposal against the ERP’s own records before anything is paid.

1. TriageHaiku sorts and scores risk
2. InvestigateEvidence from the ERP and the policy
3. ProposeResolution and vendor email
4. CheckPolicy engine checks ERP facts
5. ApproveA person signs off on large payments

• github.com/Safiullah-Rahu/matchpoint (source code on GitHub, opens in a new tab)
• The full case on neurosmith.co

Open source · Nov 2026

Sauti

Private dictation for any app, with every recording kept on the organisation’s own server.

Input: Speech in English, Arabic, Swahili or Somali

Stack: FastAPI · faster-whisper · llama.cpp · PySide6 · Kotlin

Play on YouTube1:20

Results

• Word error on the FLEURS dev set: 4.6% in English, 10.1% in Arabic, 35.3% in Swahili.
• Load test with one, three and five people dictating at once: 41 of 41 dictations completed. On a busy workstation CPU the median dictation took 21 to 28 seconds, which is why the project recommends a GPU.
• Language-model cleanup improved English only, so it is switched on for English only.

Limits

• Somali needs a dedicated speech model: Whisper’s Somali output was unusable in our tests.
• GPU support is tested in code but not yet measured on a GPU.
• The macOS and Android builds have not yet been tried on real devices.

Hold a key, speak, and cleaned-up text appears where the cursor is, on Windows, macOS or Android. Speech recognition, cleanup and translation run on the organisation’s own Proxmox server.

1. CaptureHold-to-talk on Windows, macOS and Android
2. Recognisefaster-whisper on the server
3. TidyLocal LLM: cleanup or translation
4. CheckMeaning check: numbers, negations, names
5. InsertText lands at the cursor

Play on YouTube2:00

• github.com/Safiullah-Rahu/sauti-dictation (source code on GitHub, opens in a new tab)
• The full case on neurosmith.co

Open source · Nov 2026

Switchyard

Agent workflows for a solar installer, with signed approvals and exactly-once delivery.

Input: Enquiries and service requests

Stack: OpenAI Agents SDK · MCP · Python · SQLite

Play on YouTube0:51

Results

• The demo suite runs 13 scenarios with 78 checks, including a worker killed mid-run; all pass.
• A live run on OpenAI models handled three enquiries; its reviewer caught a missing figure, and the model ignored an injected instruction.

Limits

• The CRM, calendar and email systems are simulated in the demo; the HubSpot and Slack adapters are tested only against mocked HTTP.
• The live mode was run once by hand, not in continuous integration, and the estimates in the demo are illustrative.

Agents take a solar installer from a new enquiry to a proposal, and answer service requests. Nothing binding reaches a customer without a person’s signed approval, and nothing is sent twice.

1. TriageRoutes enquiries and service requests
2. QualifyLead qualifier and estimator
3. DraftProposal and email writer, then a reviewer
4. ApproveSigned approval for binding steps
5. DeliverOutbox: sent once, then read back

• github.com/Safiullah-Rahu/switchyard (source code on GitHub, opens in a new tab)
• The full case on neurosmith.co

Open source · Nov 2026

PromptToInvoice

AI invoicing for freelancers and small businesses, where a sentence about the work becomes a draft invoice.

Input: A sentence describing the work

Stack: Next.js · Claude · PostgreSQL · Drizzle ORM · Vercel

Play on YouTube1:09

Results

• 24 unit tests cover the invoice arithmetic, the offline parser and the invite gate.
• Deployed on Vercel with Supabase as an invite-only private preview.

Limits

• Stripe billing is not built: plan changes are a demo switch for now.
• Team seats, GDPR export and delete tools, and an audit log are not built.
• Email delivery needs a Resend key and a verified domain; without them, emails are written to the server log.
• There is one invoice template.

Claude reads the sentence and fills in the client, line items, currency, tax and due date. Subtotals, discounts and tax are always calculated in code, in whole cents, never by the model.

1. DescribeA sentence about the work
2. ExtractClaude: structured outputsOffline parser as a fallback
3. CalculateTotals in code, in whole cents
4. EditEditable draft, live totals
5. SendPDF, client link and email

• github.com/Safiullah-Rahu/prompttoinvoice (source code on GitHub, opens in a new tab)
• The full case on neurosmith.co

Project videos load from YouTube only when you press play.

1. 01InsurTech · US roofing contractor·Delivered · Sep 2026Anonymised

   Autonomous claims-negotiation engine

   It reads the insurer’s estimate, finds every missing line, writes the argument — and keeps negotiating until someone answers.

   Docling extracts the PDF; a roofing-scope ontology flags omitted items and quantity and scope variances; Gemini 2.5 Flash reasons over them with safety post-processing; a PDF variance report is written; then a stateful email worker sends the argument, replies safely and follows up on its own — an IMAP state machine with recoverable jobs, AES-256-GCM-encrypted mailboxes, multi-tenant.

   Next.js 16 · FastAPI · PostgreSQL · AWS · Gemini 2.5 Flash · Docling

   1. IngestDocling PDF extraction
   2. UnderstandRoofing-scope ontologyVariance detection · omitted items · quantity & scope variances
   3. ReasonGemini 2.5 FlashSafety post-processing
   4. ActPDF variance reportStateful email worker · sends · replies safely · follows up
   5. RunRecoverable jobsIMAP state machineAES-256-GCM · encrypted mailboxesMulti-tenant

2. 02InsurTech & roofing QA·Delivered · Sep 2026Anonymised

   Computer-vision roof inspection QA

   Four photos, one verdict — an inspection that has to be taken on the roof.

   Guided four-view capture at each installation stage — decking, underlayment, valley liner, covering — inside a 250 m geofence. Gemini 2.5 Flash returns a multimodal verdict with confidence and a summary; branded PDF reports; role-based and multi-tenant.

   Gemini 2.5 Flash

   1. CaptureDecking · 4 viewsUnder-layment · 4 viewsValley liner · 4 viewsCovering · 4 views
   2. Verify250 m geofence
   3. JudgeGemini 2.5 Flash · multimodal verdictConfidence · + summary
   4. ReportBranded PDF
   5. RunRole-basedMulti-tenant

3. 03Fashion e-commerce·Handed overAnonymised

   AI fitting room: virtual try-on → 360° video

   One photo in — a 360° video of you wearing the look out.

   Kling Kolors virtual try-on, then image-to-video turns the try-on into a rotating 360° showcase. A React + three.js storefront, a Flutter app, a NestJS / AWS Cognito backend and a FastAPI AI service.

   React · three.js · Flutter · NestJS · FastAPI

   1. Try onKling Kolors · virtual try-on
   2. AnimateImage-to-videoRotating 360° showcase
   3. ShowReact + three.js storefront
   4. CarryFlutter app
   5. RunNestJS · AWS Cognito backendFastAPI AI service

4. 04Travel concierge · for a US-based tourist guide·In progress · substantially builtPublic

   Virtual Michael

   Thirty years of one expert’s judgment, turned into an avatar you can talk to.

   A real-time talking-avatar concierge: a LiveKit voice agent (Deepgram STT, ElevenLabs / Cartesia TTS, VAD, turn detection), Hedra / Anam avatars, and LlamaIndex RAG over a hand-engineered theme-park knowledge base — 23 RAG modules, 32 data sheets, 15 reference files — in a Flutter app.

   LiveKit · LlamaIndex · Flutter

   1. ListenLiveKit voice agent · VAD · turn detectionDeepgram STT
   2. KnowTheme-park knowledge base · 32 data sheets · 15 reference filesLlamaIndex RAG · 23 RAG modules
   3. SpeakElevenLabs / Cartesia TTS
   4. AppearHedra / Anam avatars
   5. CarryFlutter app

5. 05K-12 EdTech · under NDA·In production · Sep 2026Anonymised

   AI learning platform for French-speaking Switzerland

   Three thousand exercises — and infinitely many more, each rewritten for the child in front of it.

   A FastAPI AI service that regenerates worksheets, generates similar exercises, auto-corrects, drafts lesson plans and adapts content for learners with dyslexia or ADHD, over ~3,000 curriculum-aligned exercises. Three repositories, ~300 commits.

   FastAPI

   1. Curriculum~3,000 exercises · curriculum-aligned
   2. GenerateRegenerated worksheetsSimilar exercises
   3. CorrectAuto-correction
   4. PlanLesson-plan drafts
   5. AdaptVersion for learners · with dyslexiaVersion for learners · with ADHD

6. 06Music EdTech · for a London-based musician·DeliveredPublic

   TuneTutor

   Lessons in, earworms out.

   Photograph your notes: OCR, a level-aware summary, LLM-written lyrics, then a generated song with karaoke lyrics.

   Flutter · Firebase · Stripe

   1. SnapPhoto of notesOCR
   2. SummarizeLevel-aware summary
   3. WriteLLM lyrics
   4. SingGenerated songKaraoke lyrics
   5. CarryFlutter app

7. 07Enterprise data · Europe·Phases 1–5 deliveredAnonymised

   Enterprise RAG with self-drawing charts

   Ask your company’s documents anything — get the answer and the chart.

   PDF, DOCX, XLSX, CSV and JSON are ingested to S3, embedded with text-embedding-3-small into per-tenant Pinecone, and answered by gpt-4o-mini. Analytical questions come back with an auto-generated chart, titled in the language you asked in. Role-based data rooms, shared chats.

   OpenAI · Pinecone · S3

   1. IngestPDF · DOCX · XLSX · CSV · JSONS3
   2. EmbedPinecone · per tenanttext-embedding-3-small
   3. Answergpt-4o-mini
   4. ChartAnswerAuto-generated chart · titled in the language you asked in
   5. ShareShared chatsRole-based data rooms

8. 08Travel journaling·9 modules in ~6 monthsAnonymised

   AI travel companion

   A travel journal that recommends the next trip — and predicts its budget.

   A geotagged journal and map, collaborative planning and trip analytics, with a FastAPI + gpt-4o-mini service for recommendations and AI trip-budget prediction.

   Flutter · Node · FastAPI · gpt-4o-mini

   1. JournalGeotagged journal · + map
   2. PlanCollaborative planning
   3. AnalyzeTrip analytics
   4. PredictFastAPI + gpt-4o-miniRecommendations · AI trip-budget prediction
   5. CarryFlutter + Node · 9 modules · ~6 months

9. 09Media automationAnonymised

   Agentic content newsroom

   A newsroom of agents, each on the model you choose.

   CrewAI crews — Planner → Writer → Editor → Translator, plus a title-optimization crew — each agent on the model you choose (GPT-4o or Claude), auto-publishing to WordPress.

   CrewAI · GPT-4o · Claude · WordPress

   1. PlanModel per agent · GPT-4o or ClaudePlanner
   2. WriteWriter
   3. EditEditor
   4. TranslateTranslator
   5. PublishAuto-publish · to WordPressTitle-optimization crew

Reliability, uncertainty & evaluation — Calibration, conformal prediction, OOD detection and explanation audits, at every layer.

5.0★★

Fiverr rating · 64 reviews

63

five-star reviews

26

clients

16

countries

8 / 8

Upwork jobs rated 5.0

Fiverr Vetted Pro — hand-picked by the Fiverr Pro team for AI Development and Software Development · Top Rated on Fiverr · on Fiverr since 2022 · typical response ≈ 1 h

• “Safiullah is super responsive and has solid communication skills which makes it easy to collaborate and manage delivery from beginning to end.”

• “Each milestone thus far has been delivered ahead of schedule and with effective communication throughout.”

• “…provided creative solutions, and were committed to addressing any issues in requested revisions.”

• “His work was professional, exceeded my expectations, and came with great documentation.”

• “Safiullah just delivers what I asked, few questions he just gets it done!”

• “This provider blew me away and went WAY above and beyond.”

I work with a small team of developers through NeuroSmith, the AI engineering studio where I’m Chief Operations Officer.

1. 01PID motor lab
2. 02Robot-arm studio
3. 03CT reconstruction
4. 04MRI k-space explorer
5. 05Smart-wheelchair navigator
6. 06EMG prosthetic hand
7. 07Neural-network playground
8. 08Neuro-evolution cars
9. 09AgentFlow: LLM-agent workflows
10. 10MetaLab
11. 11Battery BMS
12. 12FallGuard
13. 13InkSign

Open the live lab

More public repositories

• 

  printable-games-generator

  18 generators · 297 tests

  An offline-first PWA with print-ready vector PDFs. LLM output is treated as untrusted: “A model that states 7 × 8 = 54 produces a shorter sheet, never a wrong answer key.”

  LiveGitHub

• 

  Doc-Web-AI-Chat

  55★

  Chat with PDFs; falls back to live web search.

  GitHub

• 

  Llama-3-RAG-Chatbot

  On-device

  Ollama Llama 3 + bge embeddings. Documents never leave the machine.

  GitHub