AI Agent Systems Engineer
Sohail Ali.
I build the harnesses that let AI agents ship real software, safely.
Agent pipelines with approval gates, voice layers for coding agents, and grounding that refuses to hallucinate, built on 8+ years of shipping .NET, mobile and enterprise systems.
> boot sohail.ali --profile=agent-systems ✓ core engine (.NET, 8+ yrs) ... online ✓ harness fleet ........... 6 systems ✓ agents on duty .............. 17+ ✓ approval gates ............ armed> ready.
// Systems
The fleet.
Agent harnesses, voice layers and grounding systems. Classified work is described by what it does, never who it was for.
Mothership
Agentic delivery harness
“Seven agents, one flight deck. Nothing ships without evidence.”
A project-independent harness for Claude Code and Cursor that turns every feature, bug and refactor into a gated, traceable flight plan, with its own console to fly it from.
- Agents
- 7
- Stages
- 6
- Skills
- 41
- Gate timeout
- 300s
Atlas
Enterprise agent harness
“An agent harness that maps an entire warehouse platform, from schema to ship.”
An agentic delivery environment for a large, multi-repo enterprise warehouse-management platform. It combines .NET, MAUI, Blazor and legacy WinForms/WCF, with a knowledge vault, Graph RAG over database schemas, and evidence-before-done testing.
- Skills
- 88+
- Ticket envs
- 54
- IDEs
- 3
- Policy
- Evidence-first
Helmlock
Harness v3, local-first kernel
“A 500-line kernel. Everything else is a plugin.”
The next generation of the harness line, designed in public. A small kernel, plugins for everything else, a delivery repo kept separate from a knowledge repo, and files only.
- Kernel
- ~500 LOC
- Database
- none
- Layers
- 3
- CLI
- hl
MicDrop
Voice layer for coding agents
“Talk to your coding agent. Hear it answer.”
A local-first voice layer for Claude Code, Cursor, Codex and Antigravity. You speak, it types into the agent's chat, and it reads the reply back to you. The agent is told nothing about MicDrop.
- Agents
- 4
- Cloud calls
- 0
- Modes
- 3
- Prompt changes
- none
Echo
Multi-tenant voice-agent engine
“Speaks only what it can prove.”
A voice-agent engine that can be embedded on any website. A team of specialist agents answers only from verbatim, approved knowledge, and refuses safely when it can't.
- Agents
- 5
- Grounding
- closed-world
- Embed
- 1 script tag
- Speech
- local
Tribunal
System-prompt grading instrument
“The model judges. The rubric decides.”
An instrument that scores LLM system prompts on six dimensions. LLM judges find the evidence, and a deterministic rubric turns that evidence into the score.
- Dimensions
- 6
- Judge calls
- 1 / dim
- Numbers from LLM
- 0
Anchor
Closed-KB grounding research
“Eighteen ways to stop an AI from making things up.”
Research and a testbed for a spoken RAG companion in a regulated pharma-onboarding setting, where a hallucinated answer is not an option.
- Techniques
- 18
- Domain
- regulated
- Output
- spoken
Genesis
Harness v1
“Where the harness idea was born.”
The first Claude Code agent workspace. Five agents, a folder of artifacts per ticket, and a human approval after every phase.
- Agents
- 5
- Approval
- every phase
// Live pipeline
Watch a ticket fly through the harness.
Seven agents, gated handoffs, and a human at the controls. Hover or tap a stage to see what it does.
// Harness lineage
Three generations, one idea: agents need structure.
// Mission log
From mobile apps to agent fleets.
Every mission built on the one before. Newest first. Open any mission to see its telemetry.
- M-05 2026 to now ACTIVE ▸
Orbit: Harness Fleet
A fleet of agent systems: a delivery harness, a voice layer, grounded voice agents, and an instrument that grades prompts.
- ▸Mothership: 7-agent gated delivery pipeline with its own console
- ▸MicDrop: talk to any coding agent, hear it answer
- ▸Echo, Tribunal, Anchor: closed-world grounding and evaluation
- ▸Helmlock: the next-generation kernel, in design
Systems 6Agents 17+ - M-04 2025 MISSION COMPLETE ▸
Ignition: The Agentic Turn
Turned a large legacy enterprise platform into one that agents can work on, and taught teams to work with them.
- ▸Atlas: full-SDLC skill library with a knowledge vault and Graph RAG
- ▸Genesis: first agent workspace with a human approval at every phase
- ▸Ran enablement workshops for cross-functional teams
Skills 88+IDEs 3 - M-03 2023 to now ACTIVE ▸
Command: Module Lead
Module Lead
Own delivery of key modules on large projects: mentoring engineers, guarding code quality, and keeping the work matched to what clients asked for.
- ▸Leading module teams end to end
- ▸Evidence-before-done E2E testing with Appium and Maestro
- ▸High-performance .NET APIs across multiple databases
- M-02 2019 to 2022 MISSION COMPLETE ▸
Foundation: The .NET Years
Software Engineer, then Senior Software Engineer
Enterprise backends and cross-platform apps: the base everything later is built on.
- ▸ASP.NET Core Web APIs with Dapper and SQL DTOs
- ▸Blazor (Server and WASM), .NET MAUI and Blazor Hybrid
- ▸SQL Server, Oracle, PostgreSQL, and deployment on IIS, Nginx and Docker
- M-01 2017 to 2019 MISSION COMPLETE ▸
Origin: Mobile Era
Software Engineer
Started out shipping cross-platform mobile apps to the Play Store and App Store.
- ▸Xamarin.Forms apps in C# and XAML
- ▸Real-time patient–doctor communication app
// Modules loaded
What's installed.
Agent Ops
Grounding Lab
Voice Deck
Core Engine
Systems
Data Vault
Launch Pad
Test Range
// Field notes
Notes from the lab.
all notes →Approval gates are a feature, not friction
Why my agent harnesses pause for a human at the right moments, and how to do it without becoming the bottleneck.
The model never picks the number
Using an LLM as a judge works only when the judge collects evidence and something deterministic does the scoring.
When Nemotron thinks it's Qwen
I asked a 4B NVIDIA model "tell me about yourself", and it confidently introduced itself as a different company's model.
// Comms
Building something agents should fly? Let's talk.
Open to AI engineering and agent-platform roles: remote, hybrid or on-site.
Delhi NCR, India · replies are human, not agent-generated