Current State
Built four times, about 5,000 tests, 50-user beta GO.
Project
SalesSidekick Legacy
SalesSidekick Legacy was the year-long full-stack SaaS build that got working, reached a 50-user beta GO, and still had to be thrown away as the main product direction.
The first monster build, finished and deliberately thrown away.
Origin storyEnterprise SaaSOwn the middleLegacy
Flagship Case Study
What SalesSidekick Legacy Taught
SalesSidekick Legacy was the original full-stack sales SaaS monolith: a huge build that worked, taught the architecture, and still had to be thrown away.
SalesSidekick Legacy is the clearest example of how I think about AI products under real pressure: start with operational pain, build deeply enough to learn what is actually hard, then have the discipline to narrow the product when the first answer is too big.
Where the pain lived
Enterprise sellers needed account context, CRM memory, call history, qualification logic, research, recommendations, and follow-through in one place.
What had to be true
The system had to reason iteratively, assemble context in parallel, carry seller memory, and show enough work that a rep could trust the output in a real deal.
Why it matters
It shows the learning curve behind the current work: generative coding, GTM workflow development, governance, testing, and the discipline to reduce a massive monolith into the current SalesSidekick.
Positioning Lens
SalesSidekick Legacy
Its value now is not that it is the current center of gravity. Its value is the lesson: Latif spent roughly a year building the hard thing, got it to the point where it could move toward beta, and learned that the sharper product was not a giant destination app.
Signature product decisions
- Built the original product as a full sales SaaS platform instead of a thin assistant wrapper.
- Kept the valuable operating logic while shelving the monolith and moving the current direction into SalesSidekick.
- Encoded sales methodology directly into the system so the product reflects judgment, not generic summarization.
- Protected the platform with enterprise-style test coverage and human-in-the-loop checkpoints on high-stakes workflows.
The Problem
What problem was worth solving?
I build things to solve a specific pain, not to decorate a portfolio. This is the pressure that made this project necessary.
Before the current SalesSidekick product, Latif tried to build the whole seller workflow as one full SaaS product: CRM memory, account history, call context, sales methodology, research, recommendations, approvals, document production, and follow-through in one platform.
The architecture became enormous: a real-time UI, iterative reasoning loops, a large AI kernel, automated pipelines, approval gates, and enough product scope to prove both the value and the mistake.
That build was too broad to remain the current product direction, but it became the biggest practical learning curve in generative coding, agent workflows, enterprise GTM systems, test discipline, and product scope control.
The Build
What I built around the problem
SalesSidekick Legacy combined a 78-module AI reasoning system, parallel context assembly from 12 data sources, human approval gates, CRM and research workflows, real-time reasoning views, document-production paths, a full React application, Azure infrastructure, and 295 automated pipelines. The current SalesSidekick product keeps the useful operating logic without making the original monolith the public product story.
Current status
SalesSidekick Legacy
SalesSidekick Legacy is the original sales SaaS platform and learning base. The current SalesSidekick brand carries the revenue-team product forward.
AI reasoning engine
78 modules
A dedicated AI system for understanding sales context, classifying what the rep needs, and assembling intelligence from multiple sources to produce recommendations.
Reliability
4,760+ tests
A serious automated test base across backend and frontend. The lesson was how much discipline a system this large demands.
Automation
295 pipelines
Automated workflows handling data ingestion, CRM synchronization, and background processing — keeping the intelligence layer current without manual intervention.
Human approval gates
Built-in for high-stakes actions
The AI pauses and requests approval before executing strategic workflows — ensuring human judgment on consequential decisions. Data access enforced with row-level security per tenant.
Reality Check
The painful lesson was more valuable than the monolith
The point is not to pretend the work was clean. This is what the project taught, where it changed shape, or where the boundary still matters.
01
What has been hard
SalesSidekick Legacy took roughly a year of development, got to the point where it worked, and still had to be thrown away as the main direction.
02
What had to change
The lesson was that the value was not a giant destination app. The value was context, memory, workflows, governance, and receipts that could travel into the tools people already use.
03
Current boundary
This page should make the architecture feel as large as it was while also making clear that the current SalesSidekick direction is narrower and cleaner because of what the legacy build taught.
Value
What changed because it existed
This is the practical value of the work. Where a project is unfinished, shelved, or still changing, the value is tied to what it actually taught.
Product evolution
Current SalesSidekick successor
The build clarified what should become a narrower branded product: approved skills, owned memory, manager rollups, receipts, and correction loops.
Market gate
50-user beta GO
The build reached a real beta decision point before the larger lesson became clear: the monolith was not the right long-term shape.
Codebase
~100K lines production + ~90K lines tests
The size was real, and so was the cost. This is the scale of the build that made the narrower current product direction obvious.
User experience
142 source files / 79 components
A full React application that shows the AI's reasoning in real time — sellers can see what the system is doing and why, building trust in the recommendations.
Sales skills deployed
3 live, 10 in development
Strategic Account Analysis, Post-Call Intelligence, and Competitive Positioning in production. Each skill produces graded output with human review before high-stakes actions.
Microsoft integration
4 enterprise entry points
Designed for Copilot, Teams, Dynamics 365, and Outlook — delivering intelligence wherever the seller already works, with Microsoft marketplace distribution planned.
Technical Layer
How the system is built
This is the implementation behind the work: the architecture choices, integrations, controls, and workflow decisions that made the project real enough to learn from.
Anthropic Claude
Claude for Work Team / Enterprise
LangGraph (78-module AI engine)
FastAPI + React
Azure Container Apps
Azure Durable Functions
Azure OpenAI / GPT portability
Application Insights
n8n (295 automated pipelines)
PostgreSQL + pgvector
Build Story
How the thinking unfolded
This is the reasoning path behind the output, not only the finished artifact.
01
The legacy build that taught the system
SalesSidekick was the first attempt to build the entire enterprise seller workflow as one full SaaS monolith, with context, workflow memory, methodology, approvals, and recommendations built into the product.
02
The scale was real
~100K lines of production code backed by 4,760+ tests. The frontend shows 1,632 tests passing. The important point is not only that it was large; it is that the size made the product lesson impossible to ignore.
03
What became current SalesSidekick
The monolith proved the value of seller context and reviewable workflow, but it also showed that the sharper product is the current SalesSidekick: a revenue-team product with approved skills, company-held memory, manager visibility, and correction loops.
04
Built from real seller pain
The product came from direct experience: too much admin, fragmented context, and tools built for management reporting instead of seller execution. Deal qualification (MEDDPICC), competitive analysis, and health scoring are built into the core logic — not bolted on as overlays.
Evidence
What this project shows
This evidence is strongest when it is tied to specific work rather than broad claims.
AI Platform Architecture
Designed a full sales SaaS monolith and 78-module AI reasoning engine that assembles sales context, works through multi-step analysis, and produces recommendations with human approval gates on high-stakes actions.
Sales Intelligence Integration
Connected ~30 tools spanning research, CRM, communication, and analytics into a single intelligence layer. The AI chooses which tools to use based on what the rep needs.
Real-Time Context Assembly
12 data sources (CRM profiles, territory data, deal history, market intelligence, conversation memory, and more) assembled in parallel in under 20 seconds — with graceful handling if any source is slow.
Enterprise Architecture
Single API entry point serves web, Copilot, Teams, and Slack. Sales skills run as independent cloud functions. Tenant isolation, performance monitoring, and configuration controls built into the production stack.
Sales Methodology Automation
MEDDPICC deal qualification, competitive analysis, and deal health scoring (0-100) are encoded directly into the AI skills. Outputs distinguish verified facts, estimates, and hypotheses.
Continue
Want to go deeper?
Ask the Latif AI Guide about the architecture, the commercial logic, or the hard lessons behind this project.