Ahmedabadlocal time, updates every minute
Kunj Shah presents · an applied AI engineer
I build the AIfeature, thenprove it works.
I'm Kunj Shah, an AI engineer in Ahmedabad. Founders and small teams hire me to ship agents, document search and full-stack AI apps in two to four weeks, with the tests that show exactly where they hold and where they break.
- Free 15-minute intro call
- Fixed price, in writing
- You own the code
EngineerKS
DevFest BarodaBooking projectsAhmedabad, India
Code merged upstream at OWASP and Microsoft. Every number on this page links to the work behind it.
Services
Five problems I get hired to fix.
Each one starts with a situation you might recognise, and ends with a shipped system you can open and check.
01 · Agents
Your agent says the task is done. Your database disagrees.
Agents that check their own work
An agent pipeline that compares every step against your app's real state before trusting it, and hands the task to a person when its confidence drops.
In the published Walkcast Lab run, 3 of 4 agent success claims were false. The harness flagged all three.
Agent claimsDatabase says
- 1click "Sign in"session created
- 2open /invoices/newdraft row inserted
- 3fill amount ₹48,000amount saved
- 4report "Invoice sent"status = draft
False success at step 4. Pinned to the step that caused it and saved as a regression test.
Illustrative run. In the published Walkcast Lab results, 3 of 4 agent success claims were false.
02 · RAG
The answer is somewhere in your documents. Nobody can find it.
Search and RAG over your own documents
Search across your files that cites a source for every answer, scored against a set of real questions so you know the recall before launch.
EngineerOS runs a 20-question retrieval test in CI. If recall drops below its floor, the build fails.
03 · Full Stack
You have a prototype. You need a product people can log into.
Full-stack AI apps, built and deployed
Database, API, interface and deployment in one build, with backup model providers so one provider's outage never takes your product offline.
17+ systems shipped, including a multi-tenant CRM where a test fails the build if any query could read another customer's data.
04 · Advisory
Your model bill grows every month and nobody can say why.
LLM cost and architecture review
A read of where the spend goes, then a router that sends easy requests to small models and saves the large ones for hard reasoning.
Complexity-based routing cut API spend 65% and added under 200ms per request.
05 · Computer Vision
Defect detection has to run in the field, on the device.
Computer vision on edge hardware
Detection models quantized to INT8 and tuned for Jetson-class hardware, so inference runs on the device in under 100ms.
98.7% accuracy at 4.2x manual inspection speed on real railway track footage.
Demo reels
Every build has a reel.
Screen recordings of the real products. Hover to pause the strip, or click a frame to read how it was built and tested.
01AOnRamp9 Step Pipeline
02AArchMind AI80% Coverage Gate In CI
03AEngineerOS10+ Daily Users
04AEstate3603 Tenant-Scope Test Gates
05ALattice112 Tools Indexed
06AOfferGuard AI<3s End to End
07ALearnAI+45% Quiz Completion
08ASmartFlow AI94% Accuracy
09ASENTINEL CLI13+ Analyzers
10AArchMind Research Agent4 Pipeline Stages
Selected work
Shipped, measured, and public.
Each case study states what was built, how it was tested and what it still gets wrong.
walkcast lab · replay gate5/5injected faults reproduced, and the fixed agent passed every replay
Agent Observability · MIT · v0.3.2
Walkcast
Recorded walkthroughs with narration and captions, plus a lab that turns failed browser-agent runs into reproducible regression tests checked against real application state.
- False Success Caught
- 75%
- Faults Reproduced
- 100%
- Test Cases
- 129
OnRamp
9 Step PipelineTurns any GitHub repo into a live onboarding program. A nine-step autopilot parses the codebase into a dependency and entity graph, routes every model call free-tier-first, then turns what it finds into issues, tasks, PRs, and senior reviews.
EngineerOS
10+ Daily UsersOne workspace for notes, tasks, projects and a knowledge graph, with semantic search and a multi-agent assistant that cites its sources.
Process
From first call to handover in two to four weeks.
Four steps, the same on every project, so you always know what happens next and what it costs.
- 0115 minutes, free
Intro call
You explain the problem and what done looks like. I ask about your data, your users and your deadline.
- 02Fixed price in INR
Written scope
Deliverables, timeline and price, quoted in writing. Nothing starts until you approve it.
- 032 to 4 weeks
Build
Code goes into your repository and your cloud from the first commit. 40% upfront, the rest on delivery.
- 042 revision rounds
Handover
Deployed, tested and logged, with a handoff doc your team can run it from.
Pricing
You know the price before we start.
Starting ranges for a first project, in INR. You get an exact, fixed quote in writing after the intro call.
Website and lead capture₹10,000–20,000
A site and lead-capture setup for a business that needs to start collecting inquiries.
Get a fixed quoteAI workflow automation₹15,000–35,000
One manual workflow your team repeats every week, handled by an agent that checks its own output.
Get a fixed quoteAI product or MVPFull build₹25,000–60,000+
A full-stack AI product from database to deployment, with agents, search or both, ready for real users.
Get a fixed quote- Free 15-minute intro call
- Code in your repo, owned by you
- Tests, logging and a handoff doc
- 2 rounds of revisions
Reviews and debugging ₹1,500/hour · Retainers ₹30,000–75,000/month
Open source
Code other engineers chose to merge.
67 pull requests merged into projects I don't own, each one reviewed by that project's maintainers.
OWASP/Agent-Security-Regression-Harness
Security regression workflow + goal-hijack API-key-extraction scenario for agent testing.
A security harness that never runs its own scenarios can regress silently. The job needs no credentials, since every check uses a recorded trace or a dry run, so it also runs on forks and external PRs.
+127 / -45 across 2 filesmicrosoft/AI-Engineering-CoachAGENTS.md worker + trust-flow documentation for the AI engineering coach.
AGENTS.md is what an agent reads before touching a codebase. Pointing it at the source material keeps the notes short and stops them drifting from the docs they summarise.
+37 across 1 file- 4,064
- in the last year
- 47 days
- 260 active days
- 16 days
- still going
- 117
- Sun, Oct 18, 2026
FAQ
Fair questions.
The things clients ask before sending the first message.
You're still a student. Why hire you?
I'm in the final year of a CS degree, with two internships, 17+ shipped systems and 60+ merged open-source pull requests behind me. Every project on this site links to its code and its tests, so you can check the work before paying for any of it. You also talk directly to the person writing the code.
Who owns the code?
You do. Everything lives in your repository, your CI and your cloud account from the first commit. There is no platform or subscription to keep paying for.
How does payment work?
Fixed-scope projects are 40% upfront and 60% on delivery. Hourly work (₹1,500/hour) is billed weekly against capped hours, and retainers (₹30,000–75,000/month) are invoiced net 7. UPI, bank transfer and Wise all work.
What if the result isn't right?
Two rounds of revisions are included in every fixed-scope project. The scope you approved in writing defines done, so we both measure the result against the same list.
Can I hire you full-time instead?
Yes. I'm open to AI engineer, ML engineer and full-stack roles, remote or relocating for senior roles. The resume and GitHub are linked at the bottom of this page.
Recent writing
All essays- Written, built and tested by
- Kunj Shah
- Filmed on location in
- Ahmedabad, Gujarat, India
- Merged upstream at
- OWASP · Microsoft
- Shot on
- LangGraph · FastAPI · Next.js · PostgreSQL
Coming next
Your AI feature, the one that actually works.
Send two or three sentences about it. I reply within one working day, and the first 15-minute call is free.
Hiring full-time instead?