Builder-educator working at the intersection of AI agents and human judgment — shipping practical, human-in-the-loop AI systems while teaching organizations and public-sector teams how to adopt them.
A serial hackathon builder: fourteen competition builds across DevLeague 2026, BuildForMsia, NVIDIA Inception, Lovable Vibeathon KL, and a Qwen agent hackathon — several of which became live products.
Live page: index.html · GitHub: github.com/dcontessa
This portfolio summarizes roughly 42 selected builds drawn from an active repository set, organized by theme, industry, and use case.
Client work is described but never named. Where a build was delivered for an organization, that organization is identified only by sector and scale — never by name — and no source, repository link, screenshot, or live URL is provided for it. Named projects on this page are my own products, open-source releases, or competition entries built under my own name.
Most repositories are private. Public repositories are linked where they exist.
Turns scattered context into structured decisions. A user drops in messy project context, selects a cognitive lens, and receives a decision dashboard: recommended strategic move, scenario simulation paths, a confidence score, rationale with an evidence trail, strategic risks, and a decide / delegate / delay / decline triage. Built on the C³ framework.
Industry: professional services, executive advisory · Use case: decision support, scenario planning
TypeScript React LLM orchestration decision intelligence
An autopilot agent for solo operators and boutique agencies drowning in messy leads from DMs, forms, and email. It classifies intent, urgency, budget signal, objections, and missing information; retrieves relevant business memory ranked by importance, confidence, freshness, and relevance within a context budget; then drafts a reply and follow-up plan. Human approval is required before any consequential sales action, and every workflow run is traceable.
Industry: SMB sales, agencies · Use case: lead qualification, agentic follow-up with HITL gates
TypeScript Node/Express Qwen Cloud agent memory human-in-the-loop
→ github.com/dcontessa/funnelops-autopilot
Benchmarks 100 Malaysian public-listed companies against a validated seven-construct governance model, explains why a company scores as it does, and produces an AI brief and board-ready memo. Evidence-first dashboard design: every score traces back to its source. Team build with Clement Teo and Sapttaruben A/L Krishnan; Lab 3 — Operational Sustainability & ESG.
Industry: capital markets, corporate governance, ESG · Use case: benchmarking, board reporting
TypeScript React governance analytics AI briefing
→ governiq.drcarolineang.com
A 90-second emergency response agent for people who have just transferred money to a scammer or exposed banking credentials. Deliberately scoped as a response orchestrator — not a scam detector, not a bank, not a police-reporting channel — it converts panic into the correct emergency action sequence before more money moves. Safety-first workflow design with a custom civic-signal design system.
Industry: financial crime, consumer protection, civic tech · Use case: crisis response orchestration
React 19 TypeScript Vite Telegram bot design system
→ scamsos.malaysiacorp.com
Know what was assured. And what was not. A metric-level ESG assurance and restatement registry for Southeast Asian companies, turning an ESG disclosure into a verifiable passport that records what was actually assured versus merely claimed. Solo submission, Solana Lab.
Industry: ESG, audit and assurance · Use case: disclosure verification, restatement tracking
TypeScript verification registry ESG data
Hackathons are where most of this portfolio's strongest work started. Building against a clock forces the two things that matter most in agent design — brutal scope discipline, and a demo that has to actually run in front of judges. Several of these outlived the weekend and became live products.
| Build | Competition | Track / Lab | What it proved |
|---|---|---|---|
| GovernIQ | DevLeague 2026 | Lab 3 — Operational Sustainability & ESG | Evidence-first scoring across 100 listed companies, traceable to source |
| ProofScope | DevLeague 2026 | Solana Lab (solo) | Metric-level assurance registry — what was assured vs. merely claimed |
| ScamSOS MY | BuildForMsia | Civic response | 90-second crisis orchestration, now live with a Telegram bot |
| FunnelOps Autopilot | Qwen agent hackathon | Track 4 — Autopilot Agent | Persistent agent memory, retrieval within context budget, HITL approval gates |
| GrassrootsIQ | NVIDIA Inception Grand Challenge 2026 | Growth intelligence | Signal → pattern → recommendation → next experiment loop |
| Clinician Copilot | Lovable Vibeathon KL 2026 (AWS Malaysia) | Healthcare · 1-day team build | Two front doors, one grounded knowledge engine |
| TraceMint | Solana Devnet demonstration | Sustainability finance | Approval-to-payment with no raw evidence on-chain |
| Proofline | DevLeague 2026 · Lab 1 | Financial reporting | Narrative claims checked against spreadsheet evidence in six hours |
| PayGuard | DevLeague 2026 · community Lab | AI & agentic | Every figure recomputed and pinned to a verbatim quote before payment |
| AuditFlow | DevLeague 2026 | Credit risk | Model extracts, arithmetic decides — no hallucinated figures |
| Impact Ledger | DevLeague 2026 | Logistics & insurance | Tamper-evident custody with evidence kept off-chain |
| AgentPay · ChainDraw · OrgLens | DevLeague 2026 | Agent infra · campaigns · HR | 402 payment gates, wallet-signed authority, time-aware org history |
The pattern across all seven: each build names what it refuses to do before it names what it does. That constraint is what makes a weekend prototype defensible enough to keep running afterwards.
DevLeague 2026 · shipped August 2026
The most current cluster, and the sharpest form of a thesis running through everything above: an agent may extract, but it may not decide. Each of these puts a deterministic gate between what a model produces and what actually happens — money moving, a loan approving, custody transferring.
Verification-gated agentic payments on Solana. An agent reads an invoice (PDF or XLSX); a deterministic engine recomputes every figure in Python and pins each number to a verbatim source quote. Only a fully verified invoice gets paid — a failed check stops the payment rather than flagging it for later. Industry: fintech, accounts payable · Use case: autonomous payment behind a hard verification gate
The model extracts. The arithmetic decides. A credit analyst spends hours retyping numbers from PDF financial statements into a lending decision. Automate that with a language model alone and it hallucinates figures — and in credit risk, a made-up number is a bad loan. AuditFlow splits the work along that exact seam: extraction is a model problem, decision is an arithmetic problem. Industry: credit risk, lending · Use case: statement spreading without hallucinated figures
Every financial claim needs a receipt. A six-hour build comparing narrative claims in financial-report PDFs against the figures in accompanying spreadsheets, classifying each claim as Supported or Uncertain with explainable evidence retained. Built to the DevLeague Lab 1 brief. Its documentation states plainly that the prototype makes no production privacy or PDPA compliance claim and must never process confidential customer statements — the limit is written down, not left implied. Industry: financial reporting, assurance · Use case: claim-to-evidence verification → github.com/dcontessa/proofline
IoT-assisted, tamper-evident chain-of-custody for high-value goods in transit — art, calibrated instruments, fragile machinery, sensitive electronics, where handling evidence decides whether custody is accepted and claims are paid. Solana carries authorised shipment state transitions, custody checkpoints, and inspection evidence; operational metadata and bulky evidence stay off-chain. The next slice connects an ESP32-S3 and IMU to the shipment workflow. Industry: logistics, insurance, fine art · Use case: chain-of-custody, claims evidence
An MVP of an AI agent purchasing three protected capabilities through an HTTP 402 payment gate — machine-to-machine commerce with an explicit paywall boundary rather than an open credential. Industry: agent infrastructure · Use case: metered agent access
A wallet-signed campaign flow — create, commit, draw, pay — where the organiser's wallet becomes their identity and every privileged action afterwards requires a signature from that same key. Industry: campaigns, prize draws · Use case: verifiable draw and payout
Time-aware organisational history assembled from scattered HR records. Answers two questions that are painful to reconstruct by hand: how a role changed over several years — title changes, redesignations, reporting-line shifts, splits, merges — and what a person's journey has been since joining. The two histories are woven, so you can see exactly where a person's move meets the structural change happening around them. Industry: HR, workforce planning · Use case: org design, succession evidence
Assessment-led advisory tooling — the instrument layer beneath training and consulting engagements.
Dual-track AI readiness assessment for individuals and organisations, grounded in the C³ framework. Evaluates AI maturity, generates personalised insight, and returns an actionable development path rather than a bare score. Industry: cross-sector, enterprise L&D · Use case: maturity diagnostics, capability roadmapping
Production-grade client onboarding assessment: secure Claude API integration, PDF document ingestion, persistent client database, and multi-consultant access control — built so a consulting team, not just one operator, can run it.
Industry: professional services · Use case: client intake, structured discovery
Claude API document processing multi-tenant access control
Business assessment and personal readiness diagnostics with AI-generated insight, dashboards, and exportable reports — designed to help founders and consultants turn fragmented business information into structured, assessment-led advice. Industry: SMB advisory, consulting · Use case: diagnostic reporting
Structured transformation-readiness instrument used to frame organisational change engagements. Industry: enterprise, public sector · Use case: change readiness
Readiness diagnostic for businesses assessing their capacity to enter export markets. Industry: trade, SME growth · Use case: market-entry assessment
Work aimed at government service delivery, citizen engagement, and frontline health — much of it doubling as training material for public-sector cohorts.
An agile government transformation and value-creation training platform for public-sector leaders. Eight self-paced milestones with progressive unlocking and "when ready" assessments instead of fixed deadlines — scheduling that bends around ministry workload. Includes AI-powered coaching, cross-ministry collaboration, and a framework builder covering KPA, KRA, KPI, OKR, and Balanced Scorecard.
Industry: government, public administration · Use case: leadership capability building
OpenAI integration learning platform strategy frameworks
An agentic chatbot workflow for public service automation, built on n8n: chat-based citizen engagement, automatic data logging to Sheets, and appointment scheduling through Google integrations. Designed as a reusable pattern for counter automation, citizen service portals, kiosk assistants, and community engagement.
Industry: government services · Use case: citizen intake, booking automation
n8n Google Workspace agentic workflow
A set of conversational assistants prototyped for public service enquiry handling, exploring how citizens navigate government information without knowing which department owns their question. Delivered for public-sector stakeholders — organisations not named. Industry: government · Use case: enquiry deflection, service navigation
Healthcare triage built in a one-day team hackathon: patients reach a WhatsApp intake bot while clinicians work a shared web dashboard, both powered by the same FAQ knowledge engine grounded in MedlinePlus, PubMed, RxNorm, and openFDA.
Industry: healthcare · Use case: patient triage, clinician decision support
TypeScript Supabase WhatsApp medical knowledge APIs RAG
A voice-enabled assistant built entirely on free-tier tooling, created as a teaching artifact for public-sector and education cohorts — proof that useful voice AI can be assembled with zero budget. Industry: education, public sector training · Use case: training artifact, voice interface demo → github.com/dcontessa/zero-cost-voicebot
Systems that watch a market, qualify demand, and move a pipeline — with a human in the loop at every consequential step.
Real-time competitive surveillance across social media, product platforms, and news, watching for product launches, pricing changes, and messaging shifts — then surfacing what actually changed rather than a raw feed. Industry: SaaS, consumer brands · Use case: market monitoring, positioning defence
A content and positioning system for building professional authority — translating expertise into consistent, platform-appropriate publishing. Iterated across several build environments. Industry: professional services, creator economy · Use case: thought-leadership operations
Omnichannel growth intelligence for grassroots operators: ingests creative assets across channels, scores their signals, surfaces patterns, and recommends the next controlled experiment — a disciplined signal → pattern → recommendation → next experiment loop. In development; implementation detail reserved. Industry: growth marketing · Use case: creative testing, experiment design
Lead capture, routing, and qualification tooling for small sales teams. Industry: SMB sales · Use case: pipeline operations
End-to-end funnel for electric commercial vehicle sales: diagnostic assessment, financial projections, email and SMS automation, calendar booking, and an analytics dashboard, with real-time lead scoring. Industry: automotive, fleet electrification · Use case: complex-consideration B2B sales
A prototype suite for business-owner succession: an exit-readiness scan and structured succession flows aimed at owner-operators approaching transition. Industry: SME advisory, corporate finance · Use case: exit planning, succession diagnostics
Builds concerned with what can be proven rather than what is asserted.
An evidence-receipt and approval-to-payment demonstration for local sustainability actions. Deliberately conservative in design: no raw evidence is stored on-chain and no token is issued. A working demonstration of verifiable approval flows, not a currency experiment.
Industry: sustainability, public finance · Use case: verifiable claims, approval-to-disbursement
Solana Devnet TypeScript wallet integration
A multi-page institutional web platform covering policy, standards, programmes, membership, insights, and news for a global standards body advancing responsible blockchain adoption. Organisation not named. Industry: standards bodies, blockchain policy · Use case: institutional web presence, membership
Audio and video transcription built on Whisper, with transcription history management and multi-format export — a working tool for professionals handling interview and session media at volume.
Industry: research, professional services · Use case: media transcription, research workflow
OpenAI Whisper API
Client engagements delivered as decision surfaces. All entries in this section are anonymised — no client names, repository links, live URLs, or screenshots. Described here only to show the capability and the class of problem.
Multi-tab analytical dashboard mapping creative-industry talent opportunity across a national market, with an AI strategic advisor grounded in a curated intelligence pack so recommendations cite the underlying evidence rather than improvising. Industry: creative industries, economic development · Use case: workforce strategy, policy analysis
Interactive dashboard visualising a property technology platform's business model, market position, technology architecture, competitive landscape, partnerships, and expansion potential across eight strategic vectors. Industry: proptech · Use case: investor and board communication, strategic review
Operational reporting surface consolidating delivery and performance metrics for a services business. Industry: professional services · Use case: operational oversight
Modular financial modelling delivered two ways: a multi-tenant SaaS platform and a companion mobile micro-SaaS. Covers DCF valuation, portfolio analytics, risk analysis, cash flow forecasting, and comparable company analysis, with template-driven three-statement and scenario modelling and flexible manual or CSV data input. Industry: finance, corporate advisory · Use case: valuation, scenario modelling
Security operations surface spanning vulnerability assessment, transport security diagnostics, threat detection, and educational intelligence workflows. Industry: cybersecurity · Use case: vulnerability triage, security training
A redesigned finance workspace demonstrating invoice capture, AI-assisted review, controlled journal posting, live reporting, and audit evidence capture — with a guided demo journey through the full cycle. Industry: finance operations, accounting · Use case: AP automation with audit trail
A judge portal and multi-stage scoring system for a regional technology awards programme: submission review, structured scoring across rounds, and judge workflow management. Industry: events, awards programmes · Use case: structured evaluation at scale
Scoped tight, finished properly, and used to show a cohort how a working product actually gets shipped. Most were built as classroom material; the final entry was freelance client delivery and is anonymised accordingly.
Class-demo dashboard tracking AI exhibitions, film and television summits, and esports competitions: responsive browsing, filtering, detail views, live status indicators, and saved items. Teaching point: data-driven UI, state management
Beauty discovery and shopping MVP with personalised product matching, browsing, product detail, and cart interactions. Teaching point: recommendation UX, commerce flows
Mobile-first shopping prototype for a clean skincare brand, built to demonstrate clear navigation and a low-friction browsing-to-checkout journey. Teaching point: mobile-first conversion design
Food truck ordering MVP with menu browsing, cart actions, live kitchen order tracking, and local weather — a student build demonstration. Teaching point: real-time order state, third-party APIs
Premium collectible storefront for anime and mini figures with responsive browsing, product detail, cart, and checkout preview. Teaching point: premium visual design, checkout UX
Interactive digital flipbooks packaging MBA coursework into browsable, web-native learning material. Delivered on a freelance basis for an education provider. Industry: higher education · Use case: curriculum as product
Further builds, prototypes, and explorations across the portfolio.
| Project | Industry | Use case |
|---|---|---|
| Sports recovery education site | Sports science, wellness | Non-diagnostic recovery guidance, consent-gated trial funnel |
| Sport Performance OS | Sports science | Athlete performance tracking |
| Weekly content planning dashboard | Marketing, education | Multi-platform content generation (client — anonymised) |
| Infographic generator | Marketing | Automated visual asset production |
| SVG generator | Design tooling | Programmatic graphics with effect options |
| YouTube Insight Pro | Creator economy | Video performance analysis |
| Resume Rocket | Careers, HR | Résumé generation and optimisation |
| PhD thesis writer | Academia | Long-form research writing support |
| TherapyFlow | Mental health | Practice workflow prototype |
| Wellnest Balance Beacon | Wellness | Wellbeing tracking prototype |
| Corporate services websites | Corporate services | Marketing site iterations (client — anonymised) |
| Training academy platform | Education | Courses, trainer profiles, readiness surveys, admin dashboard (client — anonymised) |
| Wellness brand site | Health and wellness | Brand and acquisition site (client — anonymised) |
| Capital Sense | Financial services | Funding access prototype |
| Finance Explorers | Financial literacy, education | Learning prototype |
| Smart Trading Solutions | Trading, fintech | Trading tooling prototype |
| Marketing SaaS | Marketing technology | Multi-tenant marketing platform |
| Data Harvest | Data operations | Collection and structuring pipeline |
| PRaivacy | Privacy, communications | Early-stage prototype |
| SoulSync · Soul Tarot | Consumer wellness | Consumer engagement prototypes |
| Personal site | — | Professional web presence |
Additionally, the public GitHub account carries an active research trail of forked work across the agent ecosystem — orchestration frameworks, subagent libraries, MCP servers, voice AI, inference engines, and spec-driven development toolkits. These are study repositories, not authored work.
- Trustworthy agents. Every agent build includes human approval gates, traceable runs, and graceful degradation — AI accountable enough to sit inside a real business decision.
- Scope discipline. The strongest builds are defined as much by what they refuse to do as what they do: a scam-response orchestrator that is explicitly not a detector; a sports recovery site that is deliberately non-diagnostic; an evidence registry that stores no raw evidence on-chain. Knowing where a system's authority ends is a design decision, not an omission.
- Evidence over assertion. Recurring pattern across governance, ESG, and advisory work — scores trace to sources, recommendations cite their grounding, and claims carry their provenance.
- Ship under pressure. Fourteen competition builds in roughly eighteen months, seven of them in a single week. Hackathons are the fastest honest feedback loop available — a judge either understands the demo in ninety seconds or the idea needs rework, and no amount of documentation rescues a build that does not run.
- Build and teach. The same systems ship as products and get packaged as training for organizations, governments, and education cohorts. The teaching artifacts are not lesser work; they are the method made visible.
Caroline Ang · builder & educator · github.com/dcontessa