About · AI Systems Architect · India

Rishabh ShuklaI architect AI systems that hold up in production

I build AI systems that are delightful to talk to and boring to depend on. By day I work as an AI Lead, architecting production AI for enterprise teams. Through Ballast, my independent studio, I turn those same patterns into whole products instead of building them for someone else.

01

How I think about architecture

THE SHORT VERSION

The model is a component, not the system

An LLM is one probabilistic component inside a system that still owes users correctness. Architecture is deciding what the model is allowed to be wrong about. Language, tone, and interpretation sit with the model. Progression, pricing, limits, and safety sit in deterministic code that can be tested and reasoned about.

Permissions belong inside retrieval

Retrieval should enforce document-level permissions during the search, not filter results after the fact. Post-filtering means the wrong documents were already retrieved, already influenced ranking, and are one bug away from reaching a user or a model context. In multi-tenant systems this is the layer where isolation actually breaks, so tenancy belongs in retrieval rather than bolted on above it.

Tenants are schema, not forks

Agent platforms should be configured through schema instead of a forked codebase per customer. Once tenant behaviour lives in data, prompts, tools, quotas, and cost ceilings become things you can inspect, diff, and roll back. Once it lives in branches, every customer is a separate migration and nothing is testable.

Retrieval quality beats model choice

Most production failures I have debugged were retrieval failures wearing a model costume: wrong chunking, missing recency, no reranking, no evaluation set. Getting grounding right moves accuracy further than upgrading the model, and it survives model changes instead of being invalidated by them.

Evaluate the system, not the demo

Evaluation should measure whether the thing works in production, not whether the demo went well or a benchmark moved. Evaluation sets, traced runs, and regression suites are what turn an AI feature into something you can change without fear. Observability on agent runs matters more than any single prompt, because prompts are the easiest thing to fix once you can see what actually happened.

Humans keep the final say

AI drafts, a person approves. Approval workflows are first-class primitives, and consent, auditability, and revocation get modelled in the data layer from day one, not bolted on later when compliance turns into a requirement. Designing the handoff between model and human is architecture work, not product polish.

Built for India first

Code-mixed input, price sensitivity, and patchy connectivity are architectural constraints, not edge cases. Systems designed for Hinglish and low bandwidth from the start behave better everywhere else too.

02

What I work on

DAY AND NIGHT

By day

AI Lead · Enterprise AI systems

I lead AI projects for enterprise teams: multi-tenant agentic platforms, permission-aware retrieval, natural-language analytics over enterprise data, and the observability and guardrail infrastructure that keeps all of it safe to operate at scale. The work is mostly architecture and the operational discipline around it, deciding where determinism belongs, how tenants stay isolated, how runs get observed, and how a system degrades when a model or a dependency fails.

Through Ballast

Founder

Arc is the current product: a fitness journal that accepts plain text in English or Hinglish and turns it into structured tracking, an adaptive plan, and honest weekly feedback, with trainers layered on top under an approval gate. Deterministic progression, fatigue, and recovery logic underneath. It is built India-first, because the details global fitness apps skip are usually where the real product opportunity is.

ArcFastAPI · Cloud Run · Supabase · Expo · Claude

Areas

What I go deep on

Systemsagentic platform architecture · multi-tenant RAG · NL-to-SQL · retrieval infrastructure · LLM observability and evaluation · prompt and context management

StackPython · FastAPI · TypeScript · Postgres/Supabase · GCP · Docker · React/Expo

In the labtrace-reflection middleware for prompt optimisation · federated learning and unlearning · edge AI on drone and IoT hardware

Building something that has to work in the real world?