Can better instructions become better visual stories?
PromptZap opened a question about taste. Chroa made the visual system legible to agents. ScriptToReel carried that thinking into a reproducible media pipeline.
PromptZap → Chroa → ScriptToReelChapter 01 · Systems builder · AI practitioner
Twenty yearsshipping systems.Still curious enoughto start again.
Banking discipline, agent harnesses, local AI, public builds, community, and the field notes that connect them.
See the chapters
Current buildAgent Field a public discussion space for builders
Live Lab shipping experiments in the open
Field Notes ideas from the workbench
View all buildsChapter 02 · Living systems atlas
The work behaves like one system.
The atlas maps how the work connects: enterprise delivery, agent tooling, local AI, product experiments, community, and public proof.
Move through the chaptersBanking SystemsCore platforms, payments, controls, and risk discipline from enterprise delivery.
Chapter 02 · The route through the work
The chapters behind the work.
Banking systems shaped the discipline. Public builds, agent harnesses, local AI, and community work keep the questions moving.
- Start
- Banking systems
and delivery - Now
- AI products
and agent harnesses - Always
- Public proof
and community
Chapter 01Banking systems2004 - 2012
Chapter 02Local AI2018 - 2021
Chapter 03Agent harnesses2021 - 2023
Chapter 04Public builds2023 - Now
Chapter 05Singapore communityAlways
Build log> ideation [02:13:07] sketching problems worth solving> research [04:42:11] user interviews, field notes> build [12:33:58] ship small, ship weekly> document [00:45:10] share the journey> still curious enough to start again.
Chapter 04 · Agent systems
MODEL
≠ AGENT.
A model can reason. A useful agent also needs tools, memory, context, permissions, evaluation and loops that make its boundaries inspectable.
MODEL 01
TOOLS 02
MEMORY 03
CONTEXT 04
PERMISSIONS 05
EVALS + LOOPS 06
THE HARNESS MATTERS.
initializing curiosity_
✓ problem discovered
✓ research collected
✓ prototype created
✓ agents connected
building_ ████████████████░░SHIP SMALL. SHIP WEEKLY.
Now / Aug 2026
Rooms in the work
Active in Singapore’s AI builder community, following questions through public sessions and industry events.
Code with Claude Singapore, AI Engineer Singapore, OpenClaw Singapore, Singapore FinTech Festival, and more. Lessons from those rooms are published with the builds they changed.
Follow the evidence ↓
Start here
AI is everywhere. A starting point is harder to find.
01 · Foundations
I am new to AI
Concepts without theatre, then one small build you can open.
- Plain-language glossary
- PromptZap
- Field notes
I want to build with agents
See how a prompt becomes a skill, a harness, and a working product.
- PromptZap
- Skillz
- Chroa
I work in enterprise
Connect experiments to provenance, permissions, and human ownership.
- Banking perspective
- Nexus Bank
- AI governance
I want local and private AI
Understand what can stay local - and when cloud models still earn their place.
- Local-first AI
- Ollama + MLX
- ScriptToReel
Curiosity trails
The projects make more sense as a sequence of questions.
Is the newest model the answer - or is the harness?
Comparing models led to a stronger conclusion: tools, permissions, context, feedback loops, and reusable skills often matter more than a leaderboard position.
Models → coding agents → skillsHow much useful AI can stay on my own machine?
Cloud routing makes comparison easy. Local inference tests privacy, control, latency, and resilience against real workflows instead of slogans.
OpenRouter → Ollama → MLX / oMLXWhat changes when agent ideas meet banking constraints?
Traceable data, bounded actions, explicit hand-offs, and clear stopping points matter when agent ideas meet banking constraints.
Systems delivery → Nexus Bank → human reviewCan a very local problem become a useful product?
A focused Singapore need became a chat-first discovery product grounded in verified public data and the way people actually ask where to makan.
Local context → verified data → HalalEats SGWhat can a game teach that a dashboard cannot?
Play makes feedback immediate. It also exposes when delightful presentation outruns the evidence underneath it.
Voice game → pet league → provenanceThe live lab
Open the things the questions became.
Every card is public evidence: an interface, source repository, or honest build story. Filter by the kind of proof you want to inspect.
Chroa
A colour-system studio that generates accessible palettes, wires them into live interface previews, and exports a DESIGN.md a coding agent can follow.
Skillz
Focused, reusable procedures that turn prompt experiments into repeatable execution.
Browse the library ↗
HalalEats SG
A chat-first halal makan companion built around plain-language questions and verified local venue data.
Try the product ↗
Nexus Bank
A public learning demo for exploring conversational assistance, financial context, and bounded agent actions. It is not a financial service.
Open the demo ↗
PromptZap
Turns rough instructions into structured prompts inside the AI tools people already use.
See PromptZap ↗
ScriptToReel
A local-first pipeline that researches, scripts, assembles, narrates, and renders short videos from a topic.
Read the source ↗Markdown Convert
One click turns a webpage into clean Markdown for agents, notes, and knowledge bases. No account or backend.
Install the utility ↗
Codex Pet League
A playful leaderboard - and a story about replacing persuasive mock data with verifiable public signals.
Read what went wrong →Logo Challenge
A fast family-friendly experiment in voice input, progressive hints, and playful feedback.
Open on GitHub ↗No public items match this view yet. Try another filter.
Continue exploringEvidence, field notes, rooms and toolkit+
Shipping signals
Learning is useful. Shipping leaves evidence.
GitHub / last year
1,289 contributions
65% commits28% pull requests7% issues
Snapshot supplied on 1 Aug 2026. The total may include private contribution activity; no private repository contents are published here.
GitHub / earned
Achievement shelf
QuickdrawPair Extraordinaire ×4Pull Shark ×3YOLO
Together they are a compact record of fast iteration, collaboration, and pull-request shipping - not a score for its own sake.
Recent credentials · 2025–2026
Tap a card to verify the credential.
Data Landscape of GenAI for Project Managers
AI delivery and data · Issued Jul 2026
Verify ↗ 2025PMIAI in Agile Delivery
Responsible AI in delivery workflows · Issued Dec 2025
Verify ↗ 2025edXIntroduction to Generative AI
Generative AI foundations · Issued Apr 2025
Verify ↗ 2025IBF SGCertified AI Risk Governance Professional
AI risk and governance · Issued Mar 2025
View ↗ 2025A-CSMAdvanced Certified ScrumMaster
Advanced agile delivery · Issued Jan 2025
Verify ↗ 2025PMIGenerative AI Overview for Project Managers
AI for project leadership · Issued Jan 2025
Verify ↗Learning in public
A map of the ideas behind the builds.
Agent systems
How tools, memory, permissions, and evaluation turn a model into a dependable workflow.
Local inference
What can stay on-device without making the product slower or less useful.
AI-readable publishing
How HTML, Markdown, JSON-LD, and clear content policies make a site easier for agents to understand.
Agent skill01
A focused, reusable procedure that gives an AI agent domain instructions, constraints, and a reliable workflow for a specific job.
See it in Skillz ↗Agent harness02
The tools, context, permissions, memory, feedback loops, and runtime around a model. The harness often decides whether model output can become dependable work.
Read “The harness matters” →Local-first AI03
AI software designed to keep models, data, or core workflows on a person’s own device whenever practical.
Open the local AI notes →RAG04
Retrieval-augmented generation: find relevant source material first, then give it to a model as context for an answer.
Data provenance05
A record of where data came from, how it changed, and which claims it can legitimately support.
See the leaderboard lesson →Vibe coding06
Building software through rapid natural-language collaboration with coding agents, with the human retaining intent, taste, verification, and ownership.
Field notes
What I learned after the demo worked.
01
02→
03→
04→
Browse every field note →
Build journey · 8 min
Three doors into one local agent.
Read the story →Build in public
Codex gave developers pets. I gave the pets a leaderboard.
Essay
What 20 years in banking taught me about shipping AI
Essay
The harness matters
SG 🇸🇬 Public event trail
Rooms That Changed the Work.
01 / 08Singapore FinTech Festival
01 · Nov 2025 · Look widerSingapore EXPO · East
AI looks different when the room speaks finance.
Twenty years in banking changes the questions I ask. I look past the demo to provenance, permissions, operational ownership, and what happens on a difficult day.
Carried intoNexus Bank, a fictional demo ↗
02 · 13 Nov 2025 · Listen closelyByteDance · Raffles Quay
The useful idea often arrives between the slides.
A conversation can reveal the missing constraint faster than another hour of polishing. The best notes are usually questions I need to test later.
Carried intoWhat banking taught me about shipping AI →
03 · 2026 · Test the edgesAWS · IOI Central Boulevard Towers
Agents become real when permissions become real.
OpenClaw, Hermes, Sarah, memory, tools and hand-offs all point to the same lesson: autonomy is useful only when its boundaries are visible.
Carried intoWhy the harness matters →
04 · 16 May 2026 · Keep showing upCapitol Theatre · City Hall
I went to learn. The paper became part of the evidence.
The memorable part was not simply appearing in a group photograph. It was being in a room where practitioners were building, questioning and learning in public.
Read the sourceOfficial MFA transcript ↗
05 · Aug 2026 · Build for someoneLorong AI · one-north
The AI should step back when two people can talk.
At a six-hour build table, 87K Windows turned memories from seniors living alone into safe, consented offers of what they could teach. No match is ever faked.
Carried intoView the 87K Windows build story ↗
06 · Aug 2026 · Spend intelligence wellCarousell Campus
Use expensive intelligence to learn, then reuse what worked.
The practical model-cost question is not which model is smartest. It is which model is reliable enough for this job after context, tools, skills, evals and retries are counted.
Read the field noteView the Value-Maxxing notes ↗
07 · 2026 · Build togetherGrab · one-north
A good workflow remembers what the team learns.
Model comparison is useful. Reusable skills, clear intent and inspectable review make the learning survive the individual session.
Carried intoThe public Skillz library ↗
08 · 2026 · Forge the workflowNational University of Singapore
The hand-offs matter more than any single prompt.
Agent systems become dependable when roles, evidence, review and failure states are designed as one workflow rather than a collection of impressive demos.
Continue exploringAgent workflow experiments →
More rooms, experiments and build notes continue on LinkedIn.
Follow Tanveer's builder journey on LinkedIn ↗Tool relationships
Not a logo cloud. A map of what each tool changes.
Research and synthesis
Get oriented broadly, then return to primary evidence.
- Genspark
- Web research
- Markdown notes
Model routing
Different collaborators for different kinds of thinking.
- OpenRouter
- DeepSeek
- Gemini models
- NVIDIA Nemotron
Coding agents
Compare planning, implementation, and review instincts.
- Codex
- Cursor
- Claude Code
- Gemini CLI
- Pi
Agent teams
Roles, hand-offs, memory, and permission boundaries.
- Hermes
- OpenClaw
- Sarah agent
- Reusable skills
Local inference
What can stay on-device without forcing every task there.
- Ollama
- MLX
- oMLX
- Local workspaces
Image and motion
Visual evidence and story material tied to real builds.
- OpenAI image models
- Flux
- Kling through Fal
Terminal and connection
Calmer interfaces for inspecting multi-agent work.
- Ghostty
- Herdr
- Tailscale
- Wispr Flow
Shipping
Useful means somebody else can open it - and see its limits.
- GitHub
- Vercel
- Railway
- Supabase
Active rotation
The models change. The curiosity does not.
OpenAI
Claude
OpenRouter
HHermes
Codex
Cursor
Gemini
Ollama
✦Genspark
Tailscale
>_Ghostty
Railway
The person behind it
Curious mind. Enterprise patience. Builder mode.
I’m Tanveer, an AI specialist in Singapore focused on agentic systems. My professional background is in delivering complex technology across global banking. My personal work is where I follow questions into agent workflows, local models, new interfaces, and useful Singapore products.
I’m active in Singapore’s AI builder community and attend public sessions and industry events, including Code with Claude Singapore, AI Engineer Singapore, OpenClaw Singapore and Singapore FinTech Festival. I bring those unfinished questions back to the work—and publish the lessons with the builds they changed.
A useful next step
Bring a question. Let’s see where it leads.
Thoughtful notes about AI agents, local-first systems, fintech, or a useful Singapore product idea are always welcome. Personal and community work only - this site does not offer paid consulting.
Singapore / 2026
STILL
BUILDING.
The evidence is public. The next question is open.