Sohaib Hamid applying to RepStack

Pakistan's #1 applicant for this exact job post*

Every RepStack hire gets an AI second brain. I already built mine.

I'm Sohaib. I'm applying for AI Implementation Strategist. I build agents that pick their own tools, remember things, ask before doing anything scary, and keep working after I've gone to bed. Below is your job description, answered line by line, with receipts.

  • $0 until you hire
  • Shortlist of 1
  • Page shipped the day after your post

*Self-ranked. n = 1. Evals pending.

01 / A small favour

Your JD has a to-do left in it. I did it for you.

repstack-jd.md
## What You'll Do
- Integrate AI with CRMs, APIs, webhooks, and
  data sources (e.g., GoHighLevel, Slack,
  Google Workspace) [confirm tool stack]

sohaib $ 

My content agent refuses to publish any page that still has a [FACT NEEDED] marker in it. It would have caught this one. No judgement. That's the job.

Models

Claude OpusClaude SonnetClaude HaikuJev (classifier)bge embeddingsFLUX (open weights)

Agents and context

Claude CodeMCP serversSkills as runbooksScheduled runnersPersistent memory

Plumbing

Cloudflare WorkersD1KVQueues + DLQDurable ObjectsCronWebhooks

Data and workspace

GmailGoogle SheetsDriveApifyDataForSEOGoogle Maps data

Not yet, said out loud

GoHighLevelSlack

No production build on either yet. Both have APIs and webhooks, which is the part I do every day. Hand me a sub-account and a week.

21,005leads found and scored in one run, for $0.70 against a $0.75 cap
20+live websites built by a pipeline that starts from a single URL
8tools on the MCP server I wrote and host myself
3agents on a schedule, one of them on the night shift

02 / Your JD, my receipts

"What you'll do" is mostly what I already did.

Left: your words, copied from the post. Right: the thing I built that matches. Where I come up short, I say so.

  1. 01

    Build agentic systems that choose tools, manage context, and complete multi-step tasks.

    Shipped

    An hourly agent on my Mac reads a queue and decides, per item, whether to enrich a lead, write a competitor report or build a whole website. It emails me for approval before anything goes out. A second agent does lead enrichment between 1 and 5 am.

  2. 02

    Implement MCP (or equivalent) to manage tools, context, and memory in production.

    Shipped

    Stash, my own second brain, runs an MCP server on a Cloudflare Worker. Eight tools: search, read, add, update, and work an action queue. Claude uses it every day. I also work daily through MCP connectors for Gmail, Sheets, Drive and Apify.

  3. 03

    Turn business requirements into scoped AI solutions: prototype → pilot → production.

    Shipped

    RFQ Foundry: plan doc on 2 November, staging build on 3 November, live on its own domain on 21 November. Stash went out in planned phases: backend and MCP, then a Chrome extension, then an Android app.

  4. 04

    Integrate AI with CRMs, APIs, webhooks, and data sources (GoHighLevel, Slack, Google Workspace).

    Mostly

    Google Workspace is wired into my agents today: they read mail, send reports and write rows to Sheets. Live sites post forms to a webhook that stores the lead and emails it. GoHighLevel and Slack are the honest gaps. Same plumbing, new logos.

  5. 05

    Build and version prompts, orchestration flows, and RAG pipelines.

    Mostly

    My prompts live as files, not in chat history: skills and runner prompts I can diff and reuse. One skill takes a URL and returns a live multi-page website. On RAG: embedding search is live in Stash. The full ask-a-question layer is specced with a 20-question eval and not shipped yet.

  6. 06

    Set up evaluation and monitoring for hallucinations, regressions, latency, and cost.

    Shipped

    My model router was tuned on 34 labelled tasks (32 right), then got 8 of 8 on a holdout set. A second model grades every answer and reruns weak ones on a stronger model. Paid runs have a hard budget cap. Pages with unverified facts cannot be published.

  7. 07

    Use AI daily in your own work: code, testing, debugging, documentation.

    Shipped

    This page was researched, written, built and deployed with Claude Code. So was nearly everything else on it. I direct, review and decide. The agents type.

  8. 08

    Document every build (SOPs/runbooks) so non-technical teams can run it.

    Shipped

    Every project carries a progress file and a README a stranger could resume from. My skills are runbooks in the literal sense: plain-English steps that either an agent or a person can follow.

  9. 09

    Work with TA, CSM, and Ops teams to find and automate high-impact workflows.

    Shipped

    My lead pipeline is a vetting process pointed at companies: find, audit, score, shortlist, enrich. Your TA team runs the same shape on people. And I spent 4 to 5 years running manufacturing operations, so I have been the ops team that software gets done to.

  10. 10

    Stay current on models, tools, and agentic patterns, and turn learnings into improvements.

    Shipped

    I route work across Haiku, Sonnet and Opus by difficulty, use a small classifier model for cheap judgement calls, and run open-weight models for embeddings and images. When a new model lands, the router gets re-tested, not re-guessed.

And the qualifications, quickly

  • Engineering degree Electrical engineering.
  • 2+ years in production 4 to 5 years running factory operations. The AI systems here are newer, all live, and all mine to keep running. Your call.
  • Real agentic systems Rows 01 and 02 above.
  • MCP, tools, memory I wrote a server. I use six more.
  • Model literacy There is a quiz below. I take it for you.
  • Hosted and open-source models Claude and Jev hosted; bge and FLUX open weights.
  • APIs, databases, real services D1, KV, queues, cron, email, webhooks.
  • Evaluate beyond accuracy Approval gates, fact markers, run logs.
  • Scaling concerns Rate limiters, dead-letter queues, caps, caching.
  • Failure modes I kept a list. It's next.
  • Small, empowered team I run a company of one plus agents.
  • English for US clients Every site linked here was written for US buyers.

03 / The Superfecta, my version

Six systems. All built with AI, all still running.

Second brain

Stash

Everything I save from LinkedIn, X, Instagram or the web lands in one library with a summary, tags and my voice note. Claude searches it through its own MCP server. You call this a Memory Vault. Same idea, already in daily use.

  • Cloudflare Worker + MCP
  • Chrome extension
  • Android app
Agents

The night shift

Three scheduled agents. One works an action queue every hour. One enriches leads between 1 and 5 am. One drafts content twice a day and waits for me to reply "publish" by email. Anything outward-facing needs my yes.

  • Lock files, logs, catch-up on wake
  • Email approvals
Cost and quality

Model Router

Splits a job into subtasks, scores how hard each one is, and sends it to the cheapest model that can do it well. A grader checks the answer and escalates the weak ones. Try it two sections down.

  • 32 of 34 on tuning set
  • 8 of 8 on holdout
Pipelines

Lead engine

4,698 map searches across 692 US cities, deduplicated into 21,005 businesses, then classified and tiered by a small model for about four cents per thousand sites. Total search bill: $0.70. The cap was $0.75.

  • Find, audit, score, enrich
  • Review queue for the maybes
URL in, site out

Website factory

One skill takes a company's URL, crawls the old site for facts and photos, rewrites the copy without inventing anything, builds the pages, scores them, deploys, and tests the form. More than 20 are live.

SaaS

Social Baddie

A multi-brand social media scheduler built in eight phases on Cloudflare's free tier. Job queue with a dead-letter queue, a rate limiter, a cron tick every minute, and 31 passing tests. Yes, that is the real name.

  • Workers, D1, KV, Queues
  • Durable Object rate limiter

04 / "Awareness of common failure modes"

Things that broke, and what I changed.

Anyone can show you the demo. Here is the incident log.

parallelism

2,000 healthy websites reported as dead

I ran a classifier across thousands of sites at 20 at a time. The network choked and the agent marked the timeouts as "unreachable".

Fix Dropped to 8 at a time, re-ran the failures, and now treat "unreachable" as a claim to verify, not a result.

state

A job stuck "in progress" forever

An agent hit a usage limit halfway through building a site and died. Its job stayed claimed, so nothing else would pick it up.

Fix Lock directories, a reset rule for orphaned jobs, and a catch-up run when the machine wakes.

hallucination

The model wanted to invent prices

Content drafts need facts only the owner knows: prices, minimums, certifications. A model will happily make them up.

Fix The agent writes [FACT NEEDED] instead of guessing, and the publish step refuses any page that still contains one.

wrong tool

The model was bad at spotting a call button

I asked a vision model whether screenshots had a phone button or stock photos. It was unreliable at small UI details.

Fix Moved those checks to plain code that reads the page. Not every question needs a model.

rate limits

Two sources started returning 429s

Enrichment from cloud IPs got rate-limited by a couple of sites, and saves came back empty.

Fix A cron retry, three attempts, then mark it thin and move on. The pipeline finishes either way.

05 / "Latency / cost / quality trade-offs"

Pick a task. Watch where I'd send it.

This replays how my router decides. The thresholds are the real ones. The tasks and their scores are examples.

Task Score0–3 Haiku Sonnet Opus
under 1.3 → Haikuunder 2.5 → Sonnetelse → Opusgrader under 0.3 → rerun one model up

Go on, pick one.

06 / Placement plan

My first 30 days, scoped before the interview.

I only know RepStack from the outside, so these are educated guesses. I'd expect the first week to replace half of them.

Week 1

Confirm the tool stack

Someone has to. Sit with TA, CSM and Ops, map where the hours actually go, and rank workflows by hours saved against the damage if the AI gets it wrong.

Week 2

One pilot, one team

Pick the dullest high-volume workflow. A likely one for TA: a first-pass screening assist that drafts the scorecard and sends the unclear cases to a person. People decide. The agent does the reading.

Week 3

Evals and a cost line

A test set built from real past cases, a grader, a regression check on every prompt change, and a per-run cost and latency number that a non-engineer can read.

Week 4

Runbook, handover, next

A one-page SOP that the team can run without me, a log of what the agent did and why, and the next two workflows already scoped. For CSM, probably an account health digest.

The human under the stack

Engineer first. Factory second. Agents now.

Electrical engineering graduate. Spent 4 to 5 years in manufacturing and still sit as a director of a textile unit, which is where I learned that a process nobody documented is a process one person owns.

Then I moved into tech and started Bronxo Tech, where I build AI systems for sales, marketing and operations. The JD says this "is not a research or model-training role". Good. I don't train models. I put them to work and stay with them after launch.

US time zones are fine. I've been keeping them anyway.

Final stage

You've tried reading CVs.
Now try structured.

Three steps. No upfront cost. No guesswork.

  1. 1Book a call. I'll bring a screen share, not slides.
  2. 2Meet your shortlist. It's me. I saved you some time.
  3. 3Hire and grow. Pay $0 until you do.

hello@sohaibhamid.com