Sheikh Muhibul Hassan

Target role · AI Engineer, Automation & Evaluation

I build AI agents that run in production, and the evaluations that prove they work.

Applied AI engineer in Kashmir, India. I author agent‑evaluation benchmarks for frontier‑model training at Handshake AI, ship LLM automation for clients in education, construction and healthcare, and measure and document the systems I build.

Email me Download resume (PDF)

01 · Role match

Three things to remember about me

A

I evaluate AI rigorously

I design containerised agent‑evaluation tasks for frontier‑model training in security/cryptography and ML‑inference domains, with deterministic verifiers and multi‑test suites.

Proof Hostile‑server end‑to‑end task merged after maintainer deep review and Pass@3 model trials. Mercor certificate in AI agent evaluation, verify ID MRC‑2026‑ZZJYYESN.

B

I ship agents and automation to production

From a real‑time Kashmiri/Urdu voice agent to an eleven‑stage revenue engine, I build LLM systems that run every day on infrastructure I operate.

Proof Multi‑agent automation that cut core operational turnaround by 40%. Self‑hosted stack (n8n, Supabase/Postgres, Docker) with watchdogs, runbooks and a written security audit.

C

I measure models and data, not vibes

I gate model outputs against ground truth and report accuracy honestly, including the numbers that look bad and why.

Proof 39,000+ photos read against a 30,998‑ID master: 78% verified coverage, zero wrong reads in audits. Kashmiri STT evaluated by word error rate, with orthography mismatch diagnosed as the main error source.

02 · Selected projects, most relevant first

What I built and how it turned out

01

Agent benchmark authoring

Handshake AI · Project Insight · 2026
Problem
Frontier models need evaluation tasks whose pass/fail result can be trusted.
What I did
Built containerised tasks with seeded, deterministic verifiers and multi‑test suites; wrote requirements so a grader can verify each criterion objectively; graded on the security property rather than brittle status codes.
Outcome
Reference oracle scores 1.0, no‑op baseline scores 0.0, static checks clean. Hostile‑server end‑to‑end task merged after maintainer deep review and Pass@3 trials.
  • Docker
  • Harbor
  • Pytest
02

Vision‑LLM data pipeline

Meva Agro · ML data programme (led)
Problem
Match 39,000+ field photos to the right record in a 30,998‑ID master list.
What I did
Four‑tier inference cascade: QR decode, RapidOCR, Qwen2.5‑VL on Colab GPU, Claude vision agents, with exact and wildcard gating against the master list, image deskewing, checkpointed uploads and SHA‑256 proofs of completeness.
Outcome
Verified coverage lifted to 78% with zero wrong reads found in audits.
  • OCR
  • Qwen2.5‑VL
  • Claude vision
  • Python
03

Kashmiri/Urdu voice agent and STT evaluation

Lumivon AI Studio
What I did
Built a real‑time voice agent (self‑hosted speech recognition, LLM, neural TTS). Benchmarked the Kashmiri STT model on a held‑out corpus and wired it into a live phone agent tested by an automated WebRTC caller.
Outcome
Separated true recognition errors from spelling‑convention mismatch, and identified orthography mismatch as the main error source. Read the write‑up.
  • STT / WER
  • Pipecat
  • TTS
04

Multi‑agent operations automation

Meva Agro
What I did
Architected multi‑agent automation with n8n, Claude Code and Codex, plus an AppSheet + Apps Script data layer for attendance, automated payroll, KPIs and approval gates.
Outcome
Core operational turnaround cut by 40%; weekly payroll decisions run from an admin dashboard.
  • n8n
  • Claude Code
  • Codex
  • AppSheet
05

Autonomous revenue engine

Lumivon AI Studio
What I did
Eleven‑stage ETL and scoring pipeline: source, enrich, validate (MX checks over DNS‑over‑HTTPS, junk‑address filters), LLM‑researched outreach and lead scoring.
Outcome
A Lead OS KPI dashboard with live Sheets export.
  • n8n
  • Supabase
  • SQL
06

LLM agents and long‑term memory

Lumivon AI Studio
What I did
Self‑hosted Hindsight agent memory, LLM routing across Claude, Groq and NVIDIA APIs under token budgets, custom Claude Code skills and MCP browser automation.
  • RAG
  • MCP
  • Groq
07

Zaproll camera and DripYap messenger

Android
What I did
Native Kotlin evidence camera with SHA‑256 manifests, PDF site reports, GPU shaders, an offline WorkManager outbox and an R8‑signed release; an end‑to‑end encrypted messenger (AES‑256 + RSA key exchange, encrypted SQLite).
  • Kotlin
  • Android
  • Crypto

03 · Experience

Where I have done it

  1. 09/2026 – present

    AI Evaluation Task Author (Freelance)

    Handshake AI · Project Insight

    Agent‑evaluation benchmarks for frontier‑model training in security/cryptography and ML‑inference domains.

  2. 2026 – present

    Founder & Lead Engineer

    Lumivon AI Studio · durritab.com

    LLM‑powered WhatsApp/voice assistants, lead engines and websites for clients in education, construction and healthcare.

  3. 2023 – present

    Software Developer

    Meva Agro Products (OPC) Pvt. Ltd.

    Led the ML data programme; multi‑agent automation (−40% turnaround); operations data layer; Zaproll camera; storage cost model for a 21‑camera, 7‑site CCTV fleet.

  4. 02/2023 – 12/2023

    Store Manager

    Nutraj

    Team of 4. Revenue +15% in year one; stockouts −25% through replenishment controls.

  5. 2020 – 2021

    KYC Executive

    Concentrix

    AML/CDD analysis, UBO discovery, PEP screening; zero compliance slippage on high‑risk escalations.

  6. 2014 – 2017

    Process Associate

    Tata Consultancy Services (TCS)

    Policy lifecycle and automated quotation dispatch for UK insurer Phoenix Life at >96% SLA accuracy.

04 · Skills

Tools I use to do the work

AI evaluation

Rubric design & grading · agent benchmark authoring · oracle / no‑op baselines · deterministic verifiers · error analysis · Claude Code · OpenAI Codex · MCP

ML & AI

Vision‑language models (Qwen2.5‑VL, Claude vision) · OCR (RapidOCR) · Whisper / Kashmiri STT · TTS · LLM APIs (Claude, Groq, NVIDIA) · RAG memory · prompt engineering · model benchmarking

Engineering

Python · Kotlin · Java · JavaScript / Node.js · REST APIs & webhooks · n8n · Docker · Linux VPS · Git · Firecrawl · Apify · CSP and rate‑limit hardening

Data & analysis

SQL (PostgreSQL / Supabase) · Google Sheets & Apps Script · AppSheet dashboards · KPI reporting · data validation & dedup · robust statistics · financial modelling

Certification

Introduction to Agents · AI Agent Evaluation

Mercor · October 2026 · recognition for rigor, calibration and clarity · Verify

Education

Master of Business Administration

Dr. APJ Abdul Kalam Technical University · 2019 · CGPA 7.26

05 · Contact

Hiring for AI engineering or evaluation?

Email is the fastest way to reach me. The resume PDF has the full detail.

muhib.basit@gmail.com Download resume LinkedIn durritab.com