AI systems & backend engineering

I build AI applications.
And the systems behind them.

Retrieval, APIs, data, and access controls.
The whole path from an idea to working software.

Explore the work GitHub
Engineering notebook
The details matter.

A smaller memory boundary.
An honest calculation.
A failure you can recover from.

Two accepted changes.
Open them up.
Interactive explanationIllustrative model · not a benchmark
dlt / SQL Server ingestion

A smaller memory boundary.

I changed dlt’s SQL Server ADBC path to ingest Parquet one row group at a time. The driver can release each buffer, while one commit per file preserves the transaction.

0 / 6 handledGroup 1 is resident. Earlier groups have been released.
You control the steps.

Six equal units are a toy model. Real row-group sizes and driver overhead vary.

Read the change dlt #3947
Read the integration boundary calculation

Toy values 2, 4, 3 and 3 at 0–3 seconds give 9 value-seconds of observed step area. The query is at 5 seconds.

Before the fix, extending the last value beyond the final sample adds 6 unsupported value-seconds, making 15 value-seconds. The corrected calculation stops at the last observed point, leaving 9 value-seconds.

This fixed toy series explains the boundary. It is not a production query result.

Read the change VictoriaMetrics #10974

Link to this explanation
Made to be inspected. Change a setting; see the decision.Pune, India

From the interface
to the infrastructure.

I work across the application and the services beneath it. Try the browser workbenches, change the inputs, and inspect the public code behind each project.

Code / Runnable examples

Runnable examples

Source, fixtures, checks, and limits for five small engineering problems.

Check an API request-body boundary

A local ASGI proof checks the body limit against a pinned source commit using fixed test doubles.

Run the body-limit proof

Remove exact repeats without hiding conflicts

Nine synthetic CSV records become seven retained records, with two exact repeats and two conflicting keys recorded in the audit.

Run the duplicate audit

Read classification counts before the score

Invented counts make precision, recall, and cutoff tradeoffs inspectable. This is arithmetic, not a model benchmark.

Run the toy count example

Synthetic example / CSV reporting

CSV Reporting Example

A synthetic campaign report from two fictional CSV files

Inspect six illustrative metrics, their formulas and a JSON calculation receipt generated from fictional CSV inputs.

No upload or live connector. This example does not establish customer results or a production reporting service.

View the synthetic CSV reporting example Read the source and documentation

Public project / AI applications

Customer Support System

An AI answer is only as useful as the system around it.

A support application connecting document ingestion, retrieval, and tool workflows with a FastAPI AI service and a TypeScript application.

The work
Streaming responses, authorization forwarding, request correlation, and readiness checks.
The stack
FastAPI, LangGraph, TypeScript
Explore the source
Read the engineering details

Retrieval connects ingestion, embeddings, and search to the answer path. The public implementation also covers JWT verification and tracing so the API boundary can be inspected alongside the AI workflow.

The browser workbench searches six project-summary passages by keyword. The actual service connects ingestion, embeddings, and retrieval; the source documents that implementation. Deployment scale and business impact are not claimed here.

Read the service documentation

Evidence checked . Source: commit 60f711e.

GitHub Actions passed on 14 Oct 2026: web/API builds and checks, fresh and repeated PostgreSQL 16 setup, Python tests, and dry fixture evaluation. Deployment, external-provider, authentication and live RAG paths remain unverified.

Merged database setup fix

CI: builds, API checks and fresh database setup

Python tests and dry fixture evaluation

Retrieval workbenchLocal keyword search

Find the passage. Choose the context.

Keyword search over six illustrative summary passages. No vector database or generated answer.
1 passage matching “authorization”0 selected
1. API1 / 1 keywords

Authorization at the API boundary

Authorization forwarding and JWT verification make the API boundary inspectable in the public implementation.

Inspected passageAuthorization at the API boundary
Service README
Context assembly 0 selected

Selected passages

Choose passages above to assemble context. Your selection stays here when you search again.

Search and selections run locally. No external data calls.

These are portfolio summaries, not production support documents. Source links open the actual service README.

Public project / Developer tools

Context Dock

Keep the idea connected to the work.

A native macOS tool for capturing tasks and matching them to project context. Built around the small moments where an idea is otherwise lost.

The work
Local agent events, decision journaling, and atomic note writes.
The stack
Swift, SwiftUI, AppKit
Explore the source
Read the engineering details

Native UI components support task capture and project selection. Local event handling and atomic writes keep the context path explicit and inspectable in source.

The browser workbench captures and exports real notes stored on this device. It is a separate interactive sketch of the native app’s workflow.

Read the setup and implementation

Evidence checked . Source: commit f20c8fb.

GitHub Actions passed shell checks, package build and tests on 14 Oct 2026 at this pull-request commit. These checks do not establish native UI behavior, external CLI compatibility or end-user runtime outcomes.

Swift package build and tests

Context workbenchBrowser sketch

Capture a task. Keep it with its project.

Browser only; your notes stay local. No native macOS session.
Task draftAdd keyboard shortcutsAdd the detail you want to keep.
Project contextContext Dock0 saved notes

Capture a task

Unsaved draft

Local notes

0 / 10

No saved notes yet. Capture your first task to put a note in its project folder.

Loading local notes…

Your browser stores up to ten notes. Download them to keep a separate copy.

Public project / Machine learning systems

Fraud Detection

From transaction to a saved explanation.

A credit-card feature scoring API with asynchronous model explanations. The prediction, queued work, and persisted result each have an explicit contract.

The work
Validated inputs, background explanations, readiness checks, and bounded input-drift monitoring.
The stack
FastAPI, Celery, Redis, PostgreSQL
Explore the source
Read the engineering details

The API accepts 30 ordered features. A Celery worker uses the same saved preprocessing bundle and persists explanation state in PostgreSQL. Pending, completed, failed, and duplicate-ID responses remain distinguishable.

Linear explanations describe feature contributions in log-odds relative to a scaled zero reference. They are not causal evidence. Input-drift monitoring checks a bounded Redis window without blocking scoring.

A separate tuned XGBoost experiment reports ROC AUC 0.973709 on a 56,962-row historical holdout. That benchmark bundle is separate from the packaged demonstration model; it does not establish current production performance.

Read the benchmark and its limits
Published container package
Transaction workbenchBrowser simulation

Send a transaction. Follow its record.

No real model, API calls, risk scores, or feature contributions.
IdleReduced motion: use Process one step.
Request0queuedPOST /predict · queued
Browser worker0runningWaiting for a job
Session store0saved0 completed / 0 failed
Simulation

Your queue is empty. Submit one job or add a batch of three.

Jobs 0/12

Independent jobs will appear here.

Selected record
The response is not the finish.

A queued or running job reads 202. Completed records read 200; saved failures read 503.

{
  "simulation": true,
  "note": "Submit a transaction to inspect it."
}
12 jobs · last 12 events

All work happens in browser memory. Completion saves a workflow receipt, not a model explanation. Reset removes only this local simulation.

Inspect the API contract

Small changes.
Real consequences.

Accepted upstream contributions to tools other engineers use. The problem, the change, and the review are all there.

11 merged pull requestsacross seven upstream projects

The complete record
prefect#22035Allow cached task results in Azure Blob Storage to be overwritten on reruns.Mergeddlt#3947Load Parquet row groups individually to limit SQL Server ingestion memory.MergedVictoriaMetrics#10974Stop integrate() from adding data beyond the end of a metric series.Mergedunsloth#5551Release a stuck IME typing state so users can send their messages.Mergedlangfuse-python#1664Support installation with wrapt 2.x.Mergedtracecat#2715Request only the Elastic alert fields a workflow needs.Mergedlollms-webui#690Clarify path sanitization and add traversal regression tests.Merged

Merged status checked October 12, 2026. Each link opens the upstream review.

How I work

The interesting part
is where it can break.

Correctness, access, and recovery are design decisions. They belong in the implementation from the start.

Make failures visible

A timeout, authorization denial, and empty result need different handling. Avoid false success.

Test the real boundary

Check database migrations on a fresh database and verify results through the API.

Design for retries

A repeated request should preserve correctness. Recovery is part of the design.

A little more context.

Professional credentials, technical practice, and experiments. Full records stay close to their sources.

Certifications & learning24 linked records across cloud, AI, data, and security

24 of 24 records

Professional certifications12

Course certificates & job simulations12

Experiments & competitive practiceKaggle notebooks, competition records, and coding profiles
LeetCode profileGeeksforGeeks profileCredly records
Tools & working knowledgeApplication code, backend services, data, and delivery

AI applications

RAG, LangGraph, LangChain, Embeddings, Reranking, Retrieval evaluation

Backend

Python, FastAPI, SQLAlchemy, Celery, Redis, REST APIs

Data

SQL, PostgreSQL, Alembic, Parquet, Data modeling

Applications

React, TypeScript, Swift, SwiftUI

Delivery

Docker, GitHub Actions, pytest, OpenTelemetry

Security

RBAC, OAuth / JWT, Tenant isolation, Credential encryption, Mutation auditing

Writing & researchFour technical articles and a co-authored research paper

Writing & research

MediumHashnode
Find me onlineCode, writing, experiments, and professional profiles

Public profiles

GitHubLinkedInKaggleMediumHashnodeCodePenLeetCodeGeeksforGeeksCodolioCredlyTIJERTwineCollectivePyPIRemoteOK

Profile sources retain their original check dates. Hashnode was user-supplied on October 17, 2026; profile content has not been independently verified.

Ashwin Upadhyay

An engineer.
All the way through.

I build AI applications and the backend systems behind them. My work moves between retrieval, APIs, data models, interfaces, and the checks that connect them.

Based in Pune, India. B.E. Information Technology, SPPU, 8.53 CGPA.

Connect on LinkedIn

Professional experience

Sylox

Feb 2026 to Present

Founding AI Engineer

Data privacy and AI backend platforms

  • Owned IRIS feature delivery across FastAPI, React/TypeScript, and PostgreSQL, from requirements through review and release.
  • Rebuilt sensitive-data classification and repaired tests, migrations, and authentication failures.
  • Built secure integrations and provider-agnostic AI verification services; proposed customer-hosted scanning architecture.

LOPhils Inc.

Oct to Nov 2025

AI Engineer · Contract

Legal-document RAG assistant

  • Built ingestion, embeddings, metadata isolation, hybrid retrieval, reranking, and grounded answers.
  • Developed asynchronous ingestion, search, and chat APIs with retries, caching, and retrieval checks.

Professional summaries; employer code is private.