Do I even need custom AI, or will an off-the-shelf tool do?

Often the off-the-shelf tool wins, and I'll tell you when it does. I only build custom when it moves a metric a tool can't, so you don't pay for a science project.

What happens when the AI is wrong?

We design for it from the start: guardrails, evals, and human fallbacks, so a 5% error rate doesn't turn into a 5-alarm fire in production.

AI Agents & Automation Engineer for Hire

Anil Pervaiz.

Hire me

10+ years client delivery
Dubai, UAE
TypeScript / Python / Next.js
Public code & live demos

AI agents & automation engineer

I work with startups and product teams to scope, build and harden AI agents, RAG and automation against real data and tools, from the first end-to-end slice through evals, observability and handover.

View selected systems Discuss a build

Hiring for an FDE? See my engineering profile →

Built in code

CiteRAG / sample-handbookhybridrerank

How long do customers have to request a refund on an annual plan?

Hybrid search
Rerank
Verify citations

Answer

Annual plans can be refunded in full within 30 days of purchase. After that, a prorated credit applies only when monthly availability fell below the SLA. Refunds go back to the original payment method in 5 to 10 business days.

3 of 3 citations verified

Product UI · sample dataHover a citation to trace it

10+ years client delivery
Dubai, UAE
TypeScript / Python / Next.js
Public code & live demos

Web & product delivery for

[01]Selected AI systems

Three builds you can inspect, each labelled for what it is.

Every card says what works today, what isn’t proven yet, and links to the code.

Open source · public repo + live demo01 / 03

CiteRAG

Citation-grounded RAG with a measurable eval loop. Hybrid search, optional rerank, per-citation verification, and a golden question set that runs against the live API.

Works today: LlamaIndex hybrid search with optional Cohere rerank, served from a FastAPI service.
Not yet: One in-memory index per process; a restart loses it unless auto-seed or an external store is used.

Case study Source Live demo

citerag.pipeline

How it runs

Question
FastAPI /query
Hybrid retrieval
LlamaIndex
Rerank
optional Cohere
Cited answer
LLM + sources
Verify citations
one pass per citation
Golden-set eval
timestamped report

In development · live demo02 / 03

SupportPilot

An embeddable AI support agent that answers from your help-center docs, escalates to humans, and writes ticket summaries.

Works today: Embeddable widget that answers from the connected help-centre content, with citations.
Not yet: No live customer traffic yet, so there are no deflection or satisfaction numbers to report.

Case study Source

supportpilot.pipeline

How it runs

User question
in-app chat
Retrieval
help-center docs
Grounded answer
with confidence
Escalate
human handoff if unsure
Ticket summary
auto-written

Hackathon build · AMD Developer Hackathon03 / 03

Autopsy

Forensic analysis of failed companies. Six specialised agents per mode, across four modes (Postmortem, Pre-Mortem, Founder Mode, Counterfactual), research in parallel, debate findings and synthesise one verdict in about 22 seconds on an AMD MI300X.

Works today: Four analysis modes, each run by six specialised agents in parallel: 24 role configurations in total.
Not yet: No benchmark of verdict quality against expert analysis; it shows orchestration, not research accuracy.

Case study Source Live demo

autopsy.pipeline

How it runs

Company + mode
4 analysis modes
Six agents
research in parallel
Debate
challenge each finding
Synthesise
one verdict
Verdict
~22s on one MI300X

All work →AI Lab: more experiments →

[02]How I work

Shipping AI that survives real users is the hard part.

Anyone can generate a demo now. The work is in the architecture, the evals, and a system your team can keep running.

Phase 01
Scope the real workflow
Define the outcome, the constraints, and what failure looks like before any code gets written.
Phase 02
Ship the risky path first
A working vertical slice against your real data and tools, in front of users, early.
Phase 03
Harden and hand over
Evals, monitoring, guardrails and documentation, then a clean transfer of ownership to your team.

A working slice first

A working vertical slice in the first two-week block. The production timeline depends on scope and integrations.

Evaluation built in

I build evaluation into AI work: a golden question set and a report you can re-run. See CiteRAG.

No lock-in

Everything I build is yours: clean code, docs and a full handover. Retainers pause anytime.

[03]Track record

AI is the current focus. Shipping for clients is not new.

Ten years of websites, products and automation for clients taught me how to scope, ship and hand over. I apply that same delivery discipline to AI systems. Earlier web & product work →

years shipping software for clients: 10+years shipping software for clients
engineers hired and led at 911 Teachers: 7engineers hired and led at 911 Teachers
bookings on the Sporty Q platform in its first year: 10K+bookings on the Sporty Q platform in its first year

[04]Pricing

Clear, simple
pricing.

Four engagement shapes. Every one starts with a fixed scope and a price you know up front.

AI Architecture SprintStart here

A fixed-scope sprint to validate, architect, and cost your AI build — a plan you own either way.

$2,000(from)

Book a 15-min call

See all engagements

What you get

Stack + data audit

System architecture diagram

Prototype direction

Model + cost recommendation

Risk + reliability notes

Costed build roadmap

EngagementInvestment

01AI Architecture Sprintfrom $2k
02Proof-of-Value Build$8k–$20k
03Production Build$25k+
04Fractional AI Leadfrom $5k/mo

[05]FAQs

Questions & answers

Something else on your mind? Ask me directly — I reply within a working day.

Ask me directly

Often the off-the-shelf tool wins, and I'll tell you when it does. I only build custom when it moves a metric a tool can't, so you don't pay for a science project.
A working vertical slice usually lands in the first two-week block. The production timeline depends on scope and integrations, and you get a dated plan after scoping.
We design for it from the start: guardrails, evals, and human fallbacks, so a 5% error rate doesn't turn into a 5-alarm fire in production.

[06]Client words

I had a website project and it was handled very gracefully. I'm happy with the work and will be doing more jobs with him. Thank you!

Tayler FreundFounder, BrandWeld

[07]Writing

Recent technical writing from the blog.

Forward Deployed Engineering

How I price a Forward Deployed Engineer engagement

How I scope and price FDE-style AI work — discovery, thin slice, harden, handoff — mapped to the published engagement tiers, without fake day rates or open-ended retainers.

August 28, 2026·8 min read

Forward Deployed Engineering

How I evaluate agent quality before handoff

A practical FDE playbook for evaluating AI agents before handoff — smoke tests, golden sets, retrieval checks, tool-call validation, guardrails, and a go-live scorecard that is not vibes on demo day.

August 26, 2026·10 min read

View all articles

Looking for an FDE or Applied AI engineer?

Engineering profile Résumé GitHub

Next step

Have an AI workflow that needs to survive contact with real users?

Tell me what’s blocked. I’ll tell you whether custom AI is warranted and what I’d prove first.

Tell me what’s blocked See engagements

Shipping AI that survives real users is the hard part.

Anyone can generate a demo now. The work is in the architecture, the evals, and a system your team can keep running.

Phase 01
Scope the real workflow
Define the outcome, the constraints, and what failure looks like before any code gets written.
Phase 02
Ship the risky path first
A working vertical slice against your real data and tools, in front of users, early.
Phase 03
Harden and hand over
Evals, monitoring, guardrails and documentation, then a clean transfer of ownership to your team.

A working slice first

A working vertical slice in the first two-week block. The production timeline depends on scope and integrations.

Evaluation built in

I build evaluation into AI work: a golden question set and a report you can re-run. See CiteRAG.

No lock-in

Everything I build is yours: clean code, docs and a full handover. Retainers pause anytime.

AI is the current focus. Shipping for clients is not new.

Ten years of websites, products and automation for clients taught me how to scope, ship and hand over. I apply that same delivery discipline to AI systems. Earlier web & product work →

years shipping software for clients: 10+years shipping software for clients
engineers hired and led at 911 Teachers: 7engineers hired and led at 911 Teachers
bookings on the Sporty Q platform in its first year: 10K+bookings on the Sporty Q platform in its first year

I turn messy AI workflows into systems your team can actually run.