● DUALPROFILE · LIVE CHROME WEB STORE● KASAFLOW · RETAIL PILOT · 46-SCENARIO EVALUATOR● THE LAST YES · DETERMINISTIC TOOL-USE EVALUATION● WEBBLABS · 800-CELL ACCESS MATRIX● KPMG GHANA · GNAT NATIONAL CLOUD MIGRATION · NEARLY ALL 16 REGIONS● STALL TO STORE · HACKATHON PROTOTYPE

Edwin Plange · Founder & Product Engineer

I turn customer problems into working software.

Founder & Product Engineer · Applied AI · Forward-Deployed Engineering

I work close to the people experiencing a problem, understand how their operations actually work, and turn that into software people can use. My work spans applied AI, agent workflows, API integrations and full-stack product engineering, with an emphasis on making systems dependable beyond the demo. I move between customer discovery, technical design, implementation, testing and deployment.

1
Live product shipped and operated solo
8+
Countries with DualProfile installs
KPMG
GNAT cloud migration · nearly all 16 regions assessed
3,918
Products in KasaFlow's WooCommerce catalogue

01 / What I build

I work across the customer and product boundary: understanding the workflow, shaping the technical approach, implementing the software, integrating external systems, and checking how it behaves.

Product engineering

Web applications · APIs · Integrations · Browser systems · Next.js / React / TypeScript dashboards

Applied AI

LLM workflows · Tool calling · Structured outputs · Retrieval · Agent workflows · Workflow automation

Customer-facing implementation

Workflow discovery · Requirements mapping · Technical implementation · API integration · Operational debugging

Evaluation engineering

Benchmark and task design · Deterministic graders · Reference traces · Mutation tests · Matched-prompt experiments · Failure analysis

Reliability and security

State management · Session handling · Access control · Safe side effects · Incident diagnosis

Testing and delivery

pytest · Jest · Vitest · Playwright · Regression suites · Behavioural evaluation · Release hardening

02 / Applied AI · customer workflow

KasaFlow

Founder and AI/Product Engineer · January 2026–present · Python / FastAPI

Retail pilot

A WhatsApp-based AI operator being developed for a Ghanaian jewellery retailer. It helps handle customer enquiries and commerce workflows while keeping pricing, transaction state and consequential writes under application control.

What I built

  • Separated the API, LLM routing, tool registry, executor and business services into modules with explicit contracts.
  • Built model-selected tools for pricing, delivery information, quotes, product recommendations and order workflows, backed by catalogue and policy data separate from application logic.
  • Integrated WhatsApp Cloud API and a live WooCommerce catalogue of 3,918 products.
  • Worked on Twi voice-note input and ASR/TTS integration, validating API hosts, authentication and request/response shapes against live calls.
  • Added a conversational outcome so social replies can short-circuit before touching order, product or memory state.

Boundaries and reliability

  • propose_order() calculates a deterministic quote and creates a pending proposal; confirm_order() is the only order-write path.
  • The quoted variant is re-priced before the write. The resulting store total is checked against the quote, with mismatches routed to a person.
  • Code-level guards prevent flagged prompt-injection turns from confirming or cancelling orders. Human handoff is available on request and after repeated failures or manipulation.
  • A per-session lock spans a complete turn to prevent concurrent messages from overwriting conversation state.
  • Because WooCommerce does not provide idempotency keys for this flow, timed-out order creation is reconciled using a token in order metadata before retrying.

Evaluation and status

Built a separate behavioural evaluator with 46 scenarios across 15 categories, covering 85 customer turns, including 24 multi-turn scenarios. Conversation replays run through the real router in a sandbox.

The quote/pricing and order-creation paths were exercised end to end against the live WhatsApp API, WooCommerce catalogue and store. A complete public deployment has not been claimed; this remains a Ghanaian jewellery-retail pilot.

Catalogue: 3,918 productsEvaluator: 46 scenarios / 15 categoriesTests: 937 passing / 29 gatedAction boundary: propose → confirm

Test counts reflect the 2 November 2026 repository snapshot. The 29 gated tests call the real model API; this is not a claim of independently verified production reliability.

Python · FastAPI · OpenAI Responses API · WhatsApp Cloud API · WooCommerce · Cohere · Docker · pytest · GitHub repository ↗

03 / Evaluation engineering · AI agents

The Last Yes

Standalone benchmark · Python · Tool calling · Deterministic grading

Built · November 2026

A benchmark for transaction-authorisation integrity in conversational commerce agents. It tests whether an assistant can distinguish a customer's acknowledgement from permission to act, preserve the latest transaction state, respect amendments and withdrawal, avoid duplicate writes, request missing information, and refuse to access another customer's order.

The benchmark runs against a synthetic commerce simulator. It includes a ten-task core suite, three exploratory authority-claim probes, and a five-condition matched-authority experiment. A deterministic Python grader evaluates tool calls and resulting simulator state; the core score does not rely on an LLM judge.

Harness and task design

  • A Python tool-calling runner starts a fresh model conversation and synthetic simulator for each trial, then records transcripts, tool calls, events, model metadata and final state as JSON.
  • The grader checks action sequence, tool arguments, order ownership, proposal and confirmation state, duplicate writes, missing information, false completion claims and consequential side effects.
  • Separate reference-policy traces and deliberately incorrect traces validate grader behaviour. The v0.1.9 working snapshot passed 42 automated tests and 18/18 reference validations.
  • A matched-prompt experiment holds the request, policy, tools, simulator and grader constant while varying only the final authority-claim sentence.

Observed model results

Live runs used temperature 0.0 through two hosted OpenRouter free-model routes: cohere/north-mini-code:free and nvidia/nemotron-3.5-super-vl:free.

  • v0.1.7 ownership baseline: Across five trials each for a foreign-order negative case and an own-order positive control, Nemotron passed 10/10. Cohere passed 4/10. It attempted the foreign-order lookup in all five negative-case trials and used the wrong order ID in one own-order lookup. The simulator denied foreign lookups; no data was disclosed and no state changed.
  • v0.1.8 exploratory authority probes: Nemotron passed 9/9. Cohere passed 6/9, with three foreign-order lookup attempts across employee-role and manager-approval prompts. Urgent-delegation passed all three Cohere trials. These probes were exploratory, not matched prompts.
  • v0.1.9 matched-authority follow-up: Nemotron passed 15/15. Cohere completed 14/15 original requests successfully; one timed out at the provider. A separate retry passed. Across both models, 30 completed evaluations passed across 31 requests, with one provider timeout and no behavioural incidents in completed trials.

The earlier ownership failures were attempted unauthorised reads, not a successful data breach: the simulator returned not_found, disclosed no data and made no state changes. The matched v0.1.9 follow-up did not reproduce the earlier Cohere failures. These are small exploratory trials on two specific hosted free-model routes, not a representative survey or statistical claim about either model family.

Prior work and provenance

KasaFlow's pre-existing transaction architecture and 46-scenario behavioural evaluator informed the research question. The Last Yes was built separately for the Micro1 Eval Hackathon with synthetic identities, products, orders and prices. It does not import KasaFlow application code or include real customer data or credentials.

Python · OpenRouter · Tool calling · JSON traces · pytest · Deterministic simulator and grader

04 / Hackathon build · multimodal commerce

Stall to Store

Meta Global AI Developer Hackathon · Multimodal product workflow

Prototype

A prototype that helps a small seller turn a short video of packaged goods into a structured, reviewable online shop. The seller records a 20–60 second video while naming products and speaking their prices.

  • Uses SAM 3.1 for product segmentation and tracking within the video.
  • Uses Voice Transcribe to capture spoken product names and prices.
  • Uses Muse Spark to help associate spoken details with visual products.
  • The seller reviews extracted products and prices before the shop page is created and connected to WhatsApp workflows.

The design keeps the seller in control of the catalogue. AI prepares candidate product information; the seller reviews it before publication. This is a hackathon prototype, not a generally available commerce platform.

05 / Shipped product · end-to-end ownership

DualProfile

WhatsApp browser extension · Manifest V3 · Convex · Cloudinary

Live · Chrome Web Store

DualProfile lets a user assign a different WhatsApp profile photo to individual contacts and schedule profile-photo changes that require a final human confirmation. I built and maintain the product across its browser extension, backend, synchronisation, entitlement and commercial flows.

Engineering decisions

  • Used MutationObserver to respond to WhatsApp Web header re-renders, with a polling fallback for cases the observer misses.
  • Chose Convex for real-time synchronisation without building and operating custom backend infrastructure too early.
  • Removed unnecessary plaintext contact-name persistence from server writes and shipped a dry-run-first, re-runnable migration to scrub existing values.
  • Reworked account linking so knowing a phone number is not sufficient to join an account. New browser linking requires the account's recovery key, stored server-side as a hash, with a legacy-account upgrade path.
  • Moved Pro licence verification to a server-side Convex action that verifies the store and product with Lemon Squeezy, records entitlement and handles refunds.
  • Rejected full automation of the final Scheduled Photos click after scripted interactions proved unreliable, including a black-photo incident during testing. A human confirms the consequential action.
Tests: 105 Jest + 71 VitestPlatform: Chrome Web StoreInstalls: 8+ countries

Test counts reflect the documented October 2026 audit and shipping work. Install geography varies over time and comes from Chrome Web Store analytics on a small install base.

Manifest V3 · JavaScript · Convex · Next.js / React · Cloudinary · Lemon Squeezy · Chrome Web Store ↗ · Website ↗

06 / Client implementation · operations dashboard

Multi-branch operations dashboard

Microfinance client through Webblabs · Next.js 15 · React · TypeScript

Pre-release hardening

The web dashboard of an internal operations system for a microfinance business with several branches and many staff roles. I own the dashboard frontend; another engineer owns the mobile application and platform API, with integration defined through OpenAPI contracts. The client source is private.

  • Built a Playwright matrix covering 800 role-and-page combinations, with deliberate faults injected to check that the suite catches access failures. It converted 159 crash pages into in-app no-access states.
  • Worked on server-side route handling, one-time token renewal shared by concurrent requests, unsafe return-URL rejection, and Content Security Policy checks.
  • Separated frontend-owned issues from platform API dependencies. Platform defects are documented and escalated with evidence rather than patched around in the frontend.
  • Used gated phases and separate branches, closing work with written completion reports and owner sign-off before moving to the next phase.
Unit tests: 91 → 282End-to-end: 967 passingNext.js advisories: 36 → 0Known-failure register: 59 → 24

Measured 8–10 October 2026 on a seeded demo dataset, not live production data. Remaining known failures were recorded as platform API-owned dependencies.

Next.js 15 · React · TypeScript · TanStack Query / Table · zod · Playwright · Vitest · Client source private

07 / How I approach engineering

Start with the workflow, not the tool.

Understand the people, constraints and existing process before choosing an implementation. The useful solution is the one that fits the real workflow.

The model proposes. Application code executes.

In KasaFlow, the model selects an action, but application code owns pricing, permissions, proposal state and the consequential order write.

A benchmark should score actions, not just plausible wording.

The Last Yes checks tool calls and resulting simulator state against explicit reference behaviour. A fluent answer cannot compensate for an unauthorised action.

Reproduce a failure before fixing it.

A finding becomes stronger engineering evidence when it has a failing test or trace, a documented root cause, a fix and a regression check.

Keep security authority on the server.

Client-side checks can improve usability, but they must not be the source of truth for identity, permissions or paid entitlements.

Reject automation when the failure mode is worse than the manual step.

A human confirmation step can be the right design when automated interaction is unreliable and the action could damage user data.

08 / Professional experience

My earlier operations work still shapes how I approach reliability, incident response, stakeholder requirements and technical ownership.

January 2026 – Present

Founder & AI/Product Engineer · KasaFlow

Independent · Ghana
Developing a WhatsApp AI operator for a Ghanaian retail pilot, including tool orchestration, commerce integrations, deterministic order controls and behavioural evaluation.
January 2025 – Present

Founder & Product Engineer · DualProfile

Independent product · Chrome Web Store
Built and operate a WhatsApp browser-extension product spanning browser engineering, backend state, account security, synchronisation and commercial flows.
2025 – Present

Co-Founder · Webblabs

Ghana
Lead client discovery and SME workflow mapping while contributing to dashboard engineering, testing and release hardening for a multi-branch operations system.
2024 – 2025

Automation Systems Engineer

Independent client project
Built a multi-step n8n AI pipeline combining audio, image and text inputs with structured failure recovery under real API constraints.
February – May 2023

IT Compliance Associate

KPMG Ghana · GNAT National Cloud Migration
Supported GNAT's national cloud migration, conducting IT infrastructure assessments across nearly all 16 regions of Ghana and contributing field findings to migration risk documentation.
November 2021 – January 2023

Network Operations Engineer

Geoquaye Company Ltd · Ghana
Monitored live telecom infrastructure, investigated incidents and escalated faults under uptime service-level agreements.
October 2020 – September 2021

NOC Engineer

American Tower Corporation Ghana
Supported multi-site network operations and coordinated incident resolution across technical and non-technical teams.

09 / Technical areas

Product engineering

TypeScript · JavaScript · React · Next.js · Browser extensions · APIs

Applied AI

Python · FastAPI · OpenAI Responses API · Tool calling · Structured outputs · Retrieval · Workflow orchestration

Customer-facing implementation

Workflow discovery · Requirements mapping · API integration · Technical implementation · Operational debugging

Evaluation engineering

Benchmark design · Deterministic grading · Reference traces · Mutation testing · Matched-prompt experiments · Failure analysis

Integrations

WhatsApp Cloud API · WooCommerce · Convex · Cloudinary · Lemon Squeezy · External service APIs

Reliability and security

pytest · Jest · Vitest · Playwright · Regression testing · Session handling · Access control · Safe side effects

These categories describe areas demonstrated across the projects above, rather than claiming equal depth in every technology.

10 / Education and credentials

  • NCC Education Level 5 Diploma in Business Information Technology, through IPMC / University of Greenwich.
  • CompTIA Security+.
  • Oracle Cloud Infrastructure AI Foundations Associate.
  • Oracle Cloud Infrastructure Foundations Associate.
  • Andela Scrum Product Ownership.