Soman Awan

The work

What I build with AI

I build with several models at once: Claude and Codex for building, a local Ollama model for bulk reading, Gemini for deep research. The skills, the permissions and the gates around them are mine.

12MCP servers wired into working pipelines
14agent roles written
89golden tests locking rendered output

The practice

A row of folder tabs in a filing tray
Layers

Tools and models

Each model does the job it is good at, inside a pipeline I control.

Building
Claude and Codex, driven through written skills and role definitions
Bulk reading
A local Ollama model, run on my own machine
Deep research
Gemini, for long literature and market sweeps
Servers
MCP servers configured and wired, one self hosted clone, none authored
At work
Work related AI use runs inside the company's own policy
A wall of small paper cards connected by drawn arrows
Layers

Agents and permissions

Permissions, refusal rules and adversarial review are part of the design.

Roles
Written role definitions with a shared orientation ritual and enforced session laws
Gated in code
The auditor's shell allows read only forensic commands
Adversarial pass
The reviewer works blind on read only database calls and discards empty reviews
State
Append only checkpoints with content hashes, exact counts and named commits
A physical stamp beside two document trays
Layers

Quality gates

The gate gets built before the output gets trusted.

Validator
Deterministic and fail closed, run over every batch before import
Second pass
Independent auditors sweep a sample of each batch, and the sample decides
Regression
Golden tests lock the rendered output per engine variant
Result
Batches split into import ready and held, on named criteria
Rows of labeled archive boxes on shelves, one pulled halfway out
Layers

The data layer

The analysis runs on plumbing I built.

Warehouse
A Postgres schema with row level security and server side functions
Pipelines
A git tracked Python codebase, crawl through parse to normalize and organize
Queries
Local SQLite mirrors carry the analysis queries
How it was built
The schema, the policies and the migrations were written with AI help and local models
A tablet, printed storyboard frames and color swatch cards in a row
Layers

Product surface and creative pipelines

The same stack ships an app, images and finished video.

App
A Flutter app whose colour, type and spacing come from tokens
Images
Generation at scale through ComfyUI under one locked style, every run logged
Video
A scripted ffmpeg pipeline assembles a finished ad and exports both aspect ratios
Edge
A deployed Cloudflare Worker runs route logic and waitlist handling
Printed wireframe sheets pinned in a row with connector strings
Layers

Design files, driven through MCP

The design work happens in the file the team reviews.

App interface
Screen design working from the token set
This site
Several design attempts, driven from the editor
Operations
End to end operations flows drawn as diagrams
Avatar tooling
Adobe AI avatar design work, methods only

Coming soon

Coming soon

n8n demonstration flows

RSS or webhook intake, API parsing and enrichment, a local Ollama summarize and route step, and a notification or storage sink. Built and run locally as demonstration flows, no production claim.