production · 2025 – present
TradePulse
An honest quant backtesting platform — a look-ahead-free event-driven engine, costs modeled by default, and a grounded NL→strategy AI copilot.
Mubin Attar · AI software engineer · Ahmedabad, India
I build production AI systems — and prove how they work.
Four live products, opened up as engineering evidence: architecture, deterministic boundaries, task-level evaluations, and the decisions behind every tradeoff. Every metric on this page links to how it was measured.
Open to AI software engineering roles — remote or Ahmedabad, India
Flagship case study
DBWhisper
A natural-language-to-SQL agent that reads your database safely — schema-aware retrieval, a fail-closed read-only validator, and multi-provider LLM fallback.
- writes to your data, ever
- 0Fail-closed read-only validator gates every query before execution; the query is rejected if it references a table outside the enrolled schema, or if no schema index exists.
- LLM providers behind one fallback
- 6OpenAI → OpenRouter → DeepSeek → Groq → Anthropic → Gemini; the first with credentials wins, Gemini's free tier is the final fallback.
- 01Retrieve schema
Only relevant tables enter context.
- 02Generate candidate
The model proposes Postgres SQL.
- 03Validate boundary
Read-only and enrolled-table rules gate it.
- 04Execute safely
A constrained connection runs accepted SQL.
- 05Measure behavior
Golden queries score the deployed path.
Selected systems
One engineering standard, applied to different kinds of uncertainty.
Each product is live. Each claim below connects to its case study and measurement method.
production · 2025 – present
CrownWager
+EV sports-betting analytics — a validated XGBoost model turned into fair-odds edge, Kelly staking, an arbitrage finder, and a graded model track-record.
production · 2025 – present
LLM Studio
A multi-user, ChatGPT-style AI platform — per-user tenancy, token streaming, multi-LLM routing, and daily quotas, Dockerized and CI'd.
How I work
Build the evidence loop into the product.
A reliable AI feature is not finished when it produces an answer. It is finished when the answer can be constrained, evaluated, and improved without guesswork.
- 01
Constrain
Define permissions, valid outputs, and failure behavior before optimizing prompts.
- 02
Measure
Test the deployed path with task-level metrics and keep excluded cases visible.
- 03
Publish
Connect claims to methods, record decisions, and expose the tradeoffs behind the result.
“The model may improvise. The boundary should not.”
Open the evaluation registryEngineering notes
The decisions are part of the work.
Short notes on evals, safety boundaries, retrieval, tenancy, and the choices that survive contact with production.
Now
What I'm working on now
Standing up evals for DBWhisper, keeping four products live, and turning this site into a running notebook — a snapshot I update by hand.
See what I'm doing nowAvailable for AI engineering roles
Need someone who can ship the model and the system around it?
I am looking for teams where AI quality is treated as an engineering problem: observable, testable, and grounded in real product behavior.
sk.mubinattar@gmail.com