I have worked in enterprise architecture for twenty years and I have built and sold seven companies along the way. I still write production code most days. Today I do two things. I work as a forward deployed engineer: I go inside a company, build the AI system it needs with its real data, ship it into the systems its staff already use, and hand it over. And I act as a fractional Chief AI Officer for companies that need someone to own the AI decisions without a full-time hire. I am not an agency, and I do not belong to one industry.
The short version
- Role
- Agentic AI architect, forward deployed engineer, fractional Chief AI Officer
- Experience
- Twenty years in enterprise architecture. Seven companies built and sold.
- Last year
- 8,202 GitHub contributions in the twelve months to 14 September 2026, across retail, trading, security, video, robotics and CPU inference. The homepage pulls that number live.
- Builds
- AI agents, multi-agent orchestration, agentic workflows, fine-tuned small language models, low-latency trading systems
- Languages
- TypeScript, Python, Rust, Solidity
- AI stack
- Claude Agent SDK, MCP servers, DSPy, GEPA, RAG, LLM fine-tuning, LLM-as-a-judge evaluation
- Platform
- Cloudflare Workers, D1, R2, Docker, Ansible, PostgreSQL, ClickHouse, Redis
- Also
- Robotics simulation and sim2real, AI safety and prompt injection detection, on-chain settlement and MEV
- Works
- Remote, worldwide. One or two engagements at a time.
- Find me
- GitHub · LinkedIn · parun@apeg.dev
What I do now
Most AI projects inside companies stop at a pilot. The demo works, the vendor leaves, and nobody inside owns the thing, so it never touches real data or real customers. The work I do is the part after the demo. I sit inside the company, usually remote, with access to its repositories, its accounts and its people. I write down the decisions that are the owner's and not mine. Then I build, in weeks, with the company's real data, and every week something ships that a real person can use through a real channel. When it runs, I write the handover as a contract and leave. The how it runs page walks through one engagement week by week, from the records.
The fractional Chief AI Officer work is the other half. Some companies do not need a system built yet. They need someone who has done this to decide what to build, what to buy, what to leave alone, and who will own it. I take that on for a few days a month. It is the same judgment as the engineering work, applied before the engineering starts.
Both are AI work. Almost everything I have shipped in the last two years has a language model somewhere in it: reading documents, planning steps, writing code, grading its own output. I use AI coding agents heavily to build, and I put the quality bar in the compiler and the test suite rather than in a document, because an agent cannot ship code that the compiler refuses.
The last twelve months
The number I point people to is on my GitHub account: 8,202 contributions in the year to 14 September 2026. That is not one project. It is a run of unrelated domains, one after another and sometimes in parallel, and I list them here because a bio should show the record rather than describe it. Client names, product names and cities are left out on purpose; the work was done inside other people's companies.
| When | What was built | What I owned |
|---|---|---|
| Sep 2024 to Mar 2026 | The operating platform for a retail chain: purchase orders, cost of goods, debit notes, stock transfers between stores, seasons and size sets, monthly targets, role-based access, and a point-of-sale storefront. Integrated into the point-of-sale, accounting and shipping systems the business already ran. Nineteen months, a team of seven, released on a version cadence. | The architecture. The design of more than 45 domain modules. The split between the transactional path and a separate analytics path so reporting never slowed the tills. The integrations into the customer's existing systems. The team. |
| Jun 2025 to Jan 2026 | A quantitative trading platform in Rust: a five-stage signal pipeline, real-time risk management, multiple exchanges, each signal running in its own isolated sub-account, a hot streaming path and a cold analytics store. | All of it, alone. Including the production defects, fixed and numbered: deadlocks, an order race, false position closure on reconnect, ghost signals. |
| Jan to Mar 2026 | On-chain systems: a real-time liquidation monitor over WebSocket streams with no polling, and an order workflow that goes from an incoming webhook through nine checks to a placed and verified order in about 210 milliseconds, measured. | All of it, alone. |
| Mar 2026 | Security: a pay-per-call firewall that protects AI agents from prompt injection, hallucinated output and stuck loops, and a smart contract audit pipeline whose prompts were rewritten by an optimizer with a language model as the judge. Real bounty submissions came out of it. | All of it, alone. |
| Jan to Sep 2026 | A video production and distribution product, built inside a founder's company: transcription, an AI planner that writes the motion graphics code itself, a render pipeline, and distribution to platforms that each have their own rules. This is the engagement the how it runs page is drawn from. | Every line of code. The decisions file. The weekly cadence. Six rounds of review corrections. The handover. |
| May to Jun 2026 | Robotics: a chat-driven operator console where a person types a task and a simulated robot carries it out, and the bridge between a physics simulator and the open robot-learning stack, which did not exist before. Published on PyPI under my name. | All of it, alone. Three releases of the bridge in twelve days. |
| Jun 2026 | CPU inference research: a 26-billion-parameter mixture-of-experts model running at 124 tokens per second on a desktop with no graphics card. Six days, a lab notebook, and a public write-up with the dead ends left in. | The measurements, the diagnosis, the quantization recipe, the write-up. |
| Jul 2026 | Teaching: a workshop where an 8-billion-parameter model went from 30 percent to 95 percent on a robot task, with a prompt written by an optimizer from the model's own failures. A model thirty times bigger scored 45 percent with the hand-written prompt. | The experiment, the deck, the notebook, and the feedback service behind it, shipped the same afternoon. |
| Aug to Sep 2026 | Three products on one platform: a webinar funnel product, an ad-acquisition product graded on a single number, and the transactional email service the other two send through. Each one runs and passes its full test suite with no vendor credentials at all. | All of it, alone. 127 unit tests and 179 endpoint tests on the webinar product, green before any account existed. |
The retail platform was a team I led. Everything else in that table was one person. I can do both, and I know which one a given problem needs.
Before that
The twenty years before this were enterprise architecture: the systems that run a business rather than the ones that demo well. Seven of the companies I built along the way were sold. Each was built the same way, with a lean team, end-to-end ownership, and a refusal to outsource anything critical. I was the one writing the trading engine, deploying the infrastructure, acquiring the customers, and debugging production at three in the morning. That is still how I work. Production trading systems, client platforms and security infrastructure live in private repositories where they belong; the public ones are listed below.
How I work
Five habits, all of them from the records of real engagements rather than from a methodology book.
- The first thing I write is not code. It is a file of the owner's decisions that I am not allowed to relitigate, and the owner's rules for touching anything live. Writing it down is what stops me arguing with it later.
- Something ships every week through a real channel. A page a real person can open, an email that actually sends, a report the owner can read. Progress nobody can touch does not count.
- The quality bar lives in the compiler. Anything that can panic or hide a bug is denied at build time. Tests, lints and type checks run before every merge. That is what makes speed safe when AI agents are writing most of the code.
- Every external service sits behind a seam with a real implementation and a mock, chosen by whether the credential exists. The whole system runs and tests before anyone gives me a key, so the customer's approvals are never the thing that blocks.
- The handover is a contract. Who it is from and to, what shipped, what was deliberately not built and why, and a numbered activation order that names who does each step. The customer's team owns it afterwards, and I mean owns.
The how it runs page has each of these with the figures and the failures.
What I am not
I am not an agency. You get me, not whoever is free. I do not sell a product or a platform, so my recommendation is not shaped by what I need to sell you. I do not write strategy decks that nobody implements; if the answer is a document, I will say so and not charge you for an engagement. And I am not a specialist in one industry. The last year covered retail, finance, security, video, robotics and inference research, and the method was the same in each: learn the rules of the business, build with its real data, ship into what it already runs, hand over.
Published work
- What is a forward deployed engineer? The job, from inside itThe role explained by someone who does it.
- Forward deployed engineering: how the engagement actually runsWeek by week, from the records of one engagement.
- What a forward deployed engineer does inside your companyThe owner's version: the first month, the decisions, the handover.
- AI in retail: what actually gets built inside a retail chainNineteen months inside a chain's operating platform.
- Forward deployed engineer vs solutions engineer vs sales engineer vs solutions architectFour roles side by side, and which one you actually need.
- Running Gemma-4 26B at 124 tokens/sec on a CPU, no GPUThe byte budget, the quantization recipe, the roofline, and the dead ends. Scripts in the public repository.
- A self-hosted LLM does not need a GPUMemory bandwidth, not VRAM, is the constraint. Measured, not assumed.
- lerobot-genesis on PyPIThe bridge between the Genesis physics simulator and Hugging Face LeRobot. There was no official link between them; this package is that link. Open source, MIT.
- Tiny model, big jobsA workshop in two files: an 8B model from 30 percent to 95 percent with a prompt written by GEPA from its own failures.
Common questions
Who is Arun Prasath E G?
An agentic AI architect and forward deployed engineer with twenty years in enterprise architecture and seven companies built and sold. I build production AI systems inside companies, with their own data and under their rules, and hand them over when they run. I also work as a fractional Chief AI Officer. The measured work is published on this site.
What does an agentic AI architect do?
Designs systems where AI models take actions rather than just answer questions: read the company's data, call its tools, follow its rules, hand results to the people and systems that need them. The architect decides what the model may read, what it must never do, who checks its work, and how it reaches the systems staff already use. Then builds it and makes sure it runs.
How is this different from hiring an AI agency?
An agency sells a team and a process, and you get whoever is assigned. I take on one or two engagements at a time and do the work myself, inside your company. I am not selling a product, so the recommendation is not shaped by what I have to sell. When it runs, I hand over and leave.
Which industries do you work in?
I do not specialize in one. In the last year the work covered a retail chain's operating platform, quantitative trading, on-chain monitoring, security research, a video product, warehouse robotics and CPU inference. The method is the same in each.
Do you work remotely?
Yes. Every engagement on this site was run remotely, inside the company's own accounts, repositories and systems, with a written weekly cadence. A decisions file and a weekly channel matter more than sitting in the same room.
How do I get in touch?
The form below asks three questions about the problem, or email parun@apeg.dev. I read every message myself and reply within two working days. If it is a fit I suggest times to talk. If it is not, I say so and why.
Arun Prasath E G
Agentic AI architect. Twenty years in enterprise architecture, seven companies built and sold, more than 8,000 commits in the last twelve months across a retail platform, an ad-acquisition platform, a transactional email service, a video product, robotics and CPU inference research. Every number on this site was measured on real hardware or taken from a real engagement's records.
Working with me
I take on one or two engagements at a time, inside the company, and I hand over when it runs. If you have a problem that sits between products, inside your own process, and you need it running rather than recommended, that is the kind of work I do.
Start here
What is a forward deployed engineer? The job, from inside itThe role explained by someone who does it: where the name came from, a normal week, the handover, and when you should not hire one.
The proof
Running Gemma-4 26B at 124 tokens/sec on a CPU, no GPUA measurement no competitor page publishes: the byte budget, the roofline, and the dead ends.