Abhishek,
Kaushik.

An engineer for businesses that haven’t met AI yet

Agent systems, retrieval, and operational AI. Six years shipping software · New Delhi.

I'm Abhishek — an AI engineer who builds agent systems that businesses can actually run.

The short version

For the last six years I've been shipping software. The past stretch has been AI and platform engineering, most recently on the agents framework behind Bluehost's AI products. Outside of that, I'm building Neev, a modular operations platform for Indian MSMEs starting with textile distribution, because the most exciting place for AI right now isn't another consumer chatbot. It's the 63 million businesses still running on WhatsApp messages and paper ledgers, 6.34 crore per the National Sample Survey Office's 2017 Key Indicators of Unincorporated Non-Agricultural Enterprises.

My way into AI was Neural Networks: Zero to Hero, Andrej Karpathy's video course on building neural networks from scratch in code. I didn't just watch it. I built micrograd and makemore from scratch to understand what I was watching. That habit, going to the foundations rather than the abstractions, is how I work on most things. Including this site.

NowBluehost · agents framework backendBuildingNeev · MSME operations platformCo-founder / CTOVeriCite · curat.moneyWritesagent systems · AI for traditional business

Abhishek Kaushik · AI engineer

The method.

In order, every time

01 · Read

The system as it actually is, and the constraints around it, before a line changes.

02 · Spec

A PRD for the shape, an ADR for every decision that would be expensive to unwind.

03 · Build

In phases, against a roadmap. Every shipped change lands in the changelog.

04 · Harden

The process gate runs before the PR opens. Done means a receipt, not a summary.

Built in the open

This site runs the process it describes. The artifacts are public.

Six years shipping software, most recently on agent systems.

Work experience

Six years of shipping software, the recent stretch of it AI and platform engineering. My day work is the agents framework backend behind Bluehost’s AI products. Around it sit Neev, VeriCite and curat.money, plus advisory work on ClusterBid. The through-line is agent-native software: modular monoliths, retrieval systems, and operational AI.

’25 —Bluehost · agents framework. Platform engineer. A hand in maintaining and continuously improving the foundational platform behind Bluehost’s agentic AI products. The specifics are under confidentiality.

’26 —Neev. Co-founder and CTO across product, architecture and build. A modular operations platform for Indian MSMEs, starting with textile distribution.

’26 —VeriCite. Co-founder and CTO, retrieval pipeline. A multi-tenant institutional RAG platform on HF TEI, Qdrant, Ory and Kubernetes.

’26 —curat.money. CTO and tech lead. A fair-comparison tool for crypto cards, built from the data pipeline up.

’26ClusterBid. Engineering advisor on process and platform. A Go and Next.js inference-cloud platform taken from monorepo checks to a real-host k3s UAT.

Selected work.

2025 — present

NeevHero case · MSME thesis VeriCiteAI systems depth Bluehost · agents frameworkProduction scale curat.moneyProduct breadth ClusterBidUAT platform engineering

Same discipline, three doors in.

Engagements

The method above is not a display piece. It is how work happens here, and it is how work with you would happen. Pick the door that matches the situation. The first step through any of them is a clarify pass and a spec you are free to reject.

S/01

Greenfield01

An agent system or product surface that does not exist yet. Changelog entries and ADRs start with the first commit, not after the demo.

EntryYou know the problem. Nobody has built the thing.

First movesA clarify interview, then a spec triad. Nothing load-bearing skips them.

You holdThe spec triad, the ADRs, a dated changelog, and a receipt for every done claim.

FitTeams with a real workflow and real data who want the process as part of the product.

S/02

Hardening02

Something already runs and it is not trustworthy yet. Observability, an eval harness, error budgets, a runbook, and the process gate wired before new features land.

EntryIt works in demos. You need it to work at 3pm on a Tuesday.

First movesRead the system as it actually is. Measure before changing a line.

You holdMeasured baselines, a wired gate, a runbook, and tests that fail when the intent breaks.

FitWhoever lost a night's sleep to their own prototype.

S/03

Adoption03

Install the regime itself: hooks, gates, the spec pipeline, loop ceilings, and an autonomy level chosen for your risk posture. Strict variants exist as shipped presets, not theory.

EntryYour team ships with agents, but nobody trusts the output.

First movesAudit what enforcement you already have. Adopt the smallest tier that bites.

You holdEnforced hooks, a spec pipeline your repos share, and decision logs you can read later.

FitOperators who want discipline that survives contact with a deadline.

Autonomy is yours to set.Ctrl

You choose how much gets decided without you, from explain-everything to decide-and-log. Secrets, destructive operations, and receipts stay enforced at every level.

Every claim on this page is checkable. Start at the ADRs and the changelog, or take the whole site as one Markdown file. ADRs, changelog, /llms-full.txt.

First-principles notes on agent systems and AI for traditional businesses.

Writing

SEP 2026Trellis 1.0: keeping the process when the coding agent changesThe stable release made Claude Code, Codex, and Pi share an explicit process contract. The rollout then showed me where installation, native behavior, and my own confidence still diverged.8 MIN READ JUL 2026GPTx: putting GPT agents inside Claude CodeClaude Code is where I want the work to run. GPTx keeps Claude in charge while selected GPT workers join through the same agent interface, inside the same project session.10 MIN READ JUL 2026Loops need brakes: five Trellis release candidates in five weeksEveryone spent this summer telling you to stop prompting and start designing loops. I agree, and I got there backwards. The brakes came first, then the map, then the gate, then a second engine.9 MIN READ

All writing

16,860 contributions in the last twelve months.

In the open

16,860 contributions · 53 weeks, oldest to newest

Counts come from the GitHub contributions API, private repositories included, and refresh daily from a scheduled job on this site’s Worker. The repositories below are the ones that job names by hand; each count is commits to that repository over the same twelve months.

SEP 2026Neev. 2,889 commits in the last twelve months.

SEP 2026VeriCite. 2,817 commits in the last twelve months.

SEP 2026curat.money. 2,109 commits in the last twelve months.

SEP 2026ClusterBid. 974 commits in the last twelve months.

Refreshed today · @Zireael26

Ship it — literally. Eat the code before the bugs eat you.

Ship It

A maze-chase where you are a blinking cursor eating code characters and four bugs hunt you with four different strategies: one follows, one ambushes ahead, one flanks off its leader, one chases only from a distance. Push a commit to frighten them. The board, motion and optional sound are generated live in the page's own pixel language — no sprites, recordings or hidden game assets.

Score: 0
Lives: 3
Characters: 0/247
State: Ready

Focus the board for arrow keys or W A S D. Swipe at least 24 pixels, or use the visible direction buttons. With motion reduced, each legal input advances one fixed step.

Caret — you, eating characters

Beetle — Direct, follows your tile

Arrow — Ambush, aims four ahead

Cross — Flank, mirrors its leader through you

Notch — Shy, chases only from afar

Commit — frightens all four bugs

Abhishek,Kaushik.