I build AI features and prove they work.

Specs before prompts, evals before launch, and an owner after it ships. For three years I've been a core engineer at Content Camel, a US B2B sales-enablement SaaS, shipping its first AI features.

A grainy print of soft hills and one snow-capped mountain under a pale blue sky

Selected work

Content Camel

An autonomous multi-agent system

Designed and built an always-on AI engineering system from scratch, as the sole engineer.

Content Camel

An autonomous content and SEO engine

Researches, writes and opens publish-ready pull requests, gated by a human-approved brief, and feeds sales a ranked list of signals.

Content Camel

The product's first AI features

An LLM pipeline over reps' real HubSpot deal conversations that surfaces which content actually closes deals, with confidence scores and evidence.

Content Camel

"Sites", owned end to end

The prospect-facing microsite product reps share with buyers. Every layer owned solo, then launched to 100% of users.

How I build

Own your copy of the data

I don't query someone else's CRM live every time I need to think. I ingest, then operate on my own copy.

1from the CRM pipeline breakdown

The model is an unreliable function

Useful, occasionally wrong, and best wrapped in the same guardrails you'd give any external call.

2from Shipping your first AI feature

Make the uncertainty a column

You're not pretending the AI is certain. You're showing your work.

3from the CRM pipeline breakdown

Parity first, then ambition

Match behaviour exactly, ship, then improve from a stable base.

4from Rebuilding a production extension

Queues are your shock absorber

A queue between producers and workers turns a spike into a steady drip.

5from Event-driven on AWS serverless

Find the tax

Write down the manual steps a person repeats more than twice a week. Not the interesting work: the glue.

6from Make a small team move like a big one
The attribution pipeline, seven stagesLeft to right: OAuth and token watch; ingest to our own copy; LLM extraction (dashed, the model); validate at the boundary (highlighted); dedup to canonical; confidence score; the report. Failures drop to a reject log.The modelDownstream: where the trust comes fromOAuth +token watch1Ingest toour own copy2LLMextraction3Validate atthe boundary4Dedup tocanonical5Confidencescore6Thereport7fails the schemaReject logvisible, countable, reviewablean unreliable functionThe LLM is one stage out of seven.
Fig. 1How I build AI features: the model is one stage of seven, and the trust comes from what surrounds it. Read the breakdown.

Writing

  1. I let an LLM read a sales team's CRM notes. Then I refused to trust it.

    Why the interesting engineering in an AI feature sits downstream of the model.

  2. Make a small team move like a big one

    The repetitive work is a hidden tax on every team. Here is how I stop paying it.

  3. Shipping your first AI feature without losing the plot

    Extraction, not magic. Treat the model like an unreliable function and wrap it accordingly.

  4. Event-driven on AWS serverless, from zero

    SQS, Lambda, DynamoDB: a mental model that scales without babysitting.

  5. Rebuilding a production browser extension

    What I'd tell myself before starting a ground-up extension rebuild.

All writing

What people say

"Hasan goes beyond merely implementing what's designed; his high-level execution and problem-solving skills lead him to explore the 'whys' and focus on the end-user experience."

Dave Shanley, Founder and CEO of Content Camel, who managed Hasan directly. More from Mark Cassar, Ayomitide Ayeni and Simon van der Weele on LinkedIn.
Mountains drawn in fine vertical engraved lines

If there's work that should be running itself, let's talk.

hasanaqeel38@gmail.com

GitHub   LinkedIn