Curiosity.
Engineered.

Designed and built a data clean room using FastAPI, PostgreSQL and Next.js, spanning onboarding, dataset management and cross-company collaboration.

Implemented versioned datasets and asynchronous DuckDB queries over S3 Parquet, supporting 34 insight templates across audience overlap, revenue, growth and demographics.

Engineered ingestion pipelines for 2–3 GB CSV datasets containing tens of millions of rows, with normalisation and Parquet conversion on AWS Batch/Fargate.

Scaled telemetry ingestion to 10K RPM with PostgreSQL, ClickHouse, and Firestore, using batching, idempotent writes, atomic aggregation, caching, and lazy span hydration.

Achieved p99 under 3 seconds with zero data loss, validated with a k6-based reconciliation harness.

Zenskar: Scaled the backend with API versioning, resource-based access control, intricate filters and ordering with optimised SQL queries across 50+ endpoints.

WeHave: Built a data clean room using FastAPI, PostgreSQL and Next.js, spanning onboarding, dataset management and cross-company collaboration.

Luminum: AI RAG chatbot over custom knowledge using FastAPI, Next.js and AWS.

Arven: Email AI agent integrating Gmail API, FastAPI and Amazon SES.

WeHave: Processed 2–3 GB CSV datasets with tens of millions of rows on AWS Batch/Fargate, with Parquet conversion, SHA-256/HMAC hashing, PII detection and deduplication.

WeHave: Integrated AI-assisted column mapping and Airbyte connections for S3, HubSpot and Salesforce, supporting 2+ downstream providers.

WeHave: Reduced API response times by 90%+ through a serverless-to-server migration.

WeHave: Built a LocalStack environment covering 7+ AWS services, including Comprehend and Clean Rooms, improving development efficiency by 90%.

Google SoC / EvalAI: Built auto-scaling EC2 workers and auto-cancel mechanisms, cutting AWS running costs by over 50%.

Limited: Cut processing time from 60s to under 5s by offloading 10+ long-running workflows to concurrent SQS and Celery workers with task progress tracking.

Limited: Moved scraping, embeddings, Linear/Slack/GitHub integrations and routine tasks onto async worker queues.

Zenskar: Built two-way Salesforce and Zoho synchronisation across 7+ objects using Celery, APIs, webhooks and Apex triggers, with CI-automated testing.

Deployed a Redis rate limiter using the token bucket algorithm, reducing outgoing requests hitting external API limits by 99.99% and keeping external services fault tolerant.

Deployed isolated GCP infra using Pulumi and CI/CD, with env-aware IaC configuration, public Gateway routing, and automated deployment and teardown.

Designed a lower-cost shared staging and dev environment and optimised cost for prod, reducing costs by 60%.

Diagnosed a runaway LLM pipeline behind sustained five-figure cloud spend and isolated 4 root causes.

Drove remediation through caps, checkpointing, backoff, cancellation, and better pricing controls.

Implemented centralised logging with OpenTelemetry and SigNoz across 10+ services, spanning EC2 servers, scripts, microservices, and AWS Lambda functions.

Google SoC / EvalAI: Built auto-scaling EC2 workers and auto-cancel mechanisms in a Python, Django and Docker stack, cutting AWS running costs by over 50%.

Google SoC / EvalAI: Bolstered code integrity with comprehensive tests, raising coverage to 85%.

Upastithi: Django attendance system used by 300+ tutors and students.

Zenskar: Integrated Salesforce and Zoho across the full stack, using Celery, APIs, webhooks, and Apex triggers for two-way synchronisation across 7+ objects.

WeHave: Integrated Airbyte connections for S3, HubSpot and Salesforce into ingestion pipelines, supporting 2+ downstream providers.

Secured multi-tenant API access with workspace-scoped tokens, SHA-256 token storage and Supabase Row-Level Security.

Unified session and Bearer-token authentication and added documented, paginated APIs across core CRM resources.

Enabled SAML SSO for 3+ enterprise customers.

Built 10+ production proof-of-concepts using Go, Kafka and other tools.

Architected 5+ end-to-end AI projects from scratch, building RAG-based chatbots, a voice-testing LLM agent, and solutions utilising GPT, Anthropic, and Vertex AI.

Luminum: AI RAG chatbot over custom knowledge using FastAPI, Next.js and AWS.

Phonos: Voice agents and LLM testing using Bland AI and Python.

EngManager: AI agent for engineering teams integrating Linear and GitHub.

Arven: Email AI agent integrating Gmail API, FastAPI and Amazon SES.

Expanded the LLM Gateway SDK to 300+ models across 10+ providers, supporting streaming, embeddings, MCP, and tool calls in a standardised format.

The SDK underpins the Platform and Proxy, reaching 600+ GitHub stars and 300K+ downloads/month.

Developed the ecosystem for end-to-end self-improving agent development and evaluation across macOS, Linux, and Windows.

Shipped CLI/SDK releases, 10+ integrations, 5+ cookbooks, 2+ skills, and support for 10+ MCPs.

Fabricated a hosted LLM Gateway Proxy at 1K RPM, supporting completion and streaming chat, embeddings, authentication, and tool calls across 9+ providers.

Handled provider-specific transformations, trace/span capture, error handling, and cost calculation.

Drove core components of a 10-stage agent self-improvement pipeline, mining intents from production traces to produce auto-generated evaluators at scale.

Built its quality validation framework utilising 3+ SOTA benchmarks, including tau², tau³, and SWE-bench.

Productionized a Claude-powered CRM copilot on AWS Bedrock, enabling PDF ingestion, multimodal attachments, persistent conversations and automated operational suggestions.

Implemented prompt caching, scheduled suggestion processing and robust attachment validation and error handling.

Built RAG-based chatbots, a voice-testing LLM agent, and solutions utilising GPT, Anthropic, and Vertex AI across 5+ AI products.

Planned a custom algorithm with HDBSCAN using embeddings and cosine similarity to cluster sentences in real time, achieving 95% precision for dashboard content analysis.

Established custom heuristics with data analysis, data cleaning, and an engineered prompt pipeline, building a framework that evolves with changing data.

Reached 95% accuracy in classifying and extracting relevant images and details from emails.

EngManager: AI agent for engineering teams integrating Linear and GitHub.

Limited: Ran Linear, Slack and GitHub integrations on async SQS and Celery workers with task progress tracking.

Arven: Email AI agent integrating Gmail API, FastAPI and Amazon SES.

Built drag-and-drop Agent workflows linking prompts, APIs, datasets, and evaluations through prompt DAGs.

Led the V1-to-V2 platform migration, dynamic dataset columns, and synchronised editing across core schemas.

Designed and delivered 7+ production systems for organisations including RSC Anderlecht, PSV, Feyenoord, Tomorrowland and Royal Antwerp FC.

Built a multi-tenant Sponsorship OS, selected by 7 Belgian Pro League clubs for 2026/27, with season-aware rights and reusable sponsorship packages.

Planned delivery across 4 views—Gantt, Kanban, list and table—with hierarchical grouping, drag-and-drop scheduling and synchronised updates.

Developed ticketing, hospitality and parking operations with season credits, partner self-service, guest management and centralised reservation validation.

Implemented adjacency-based smart seating with locked-seat constraints and batch assignment.

Extended a 3D venue editor with multi-object transforms, snapping and layout tracing.

Shipped a gamified, quest-based onboarding flow modelled as a multi-edge directed graph with dynamically branching completion paths of 30+ tasks.

Featured on Mobbin, it drove a 2× increase in signups and improved long-term retention by 132%.

Built drag-and-drop Agent workflows linking prompts, APIs, datasets, and evaluations through prompt DAGs.

Inbox management app built with React Native, React, FastAPI and Celery.

Built Audience Insights using Meta's Reach Estimate API and page scrapers capturing likes, comments and engagement signals.

Enabled reach, affinity and overlap analysis with responsive Chart.js visualisations, persistent annotations and CSV validation, from Figma designs.

Built first-party audience activation across Meta and TikTok, enabling custom-audience creation, synchronisation and match-rate reporting.

Delivered performance dashboards across 3 advertising platforms—Meta, Google Ads and TikTok—to track campaign metrics and activation effectiveness.

Built Audience Insights using Meta's Reach Estimate API and page scrapers capturing likes, comments and engagement signals.

Sampark: Peer-to-peer video chat using WebRTC, Socket.IO, Node.js and React, built at the NIT-Rourkela Hackathon.

Saarthi: Education platform using YouTube API, MERN, Firebase, WebRTC and Socket.IO; a finalist at NIT-R.

Pahchan: Portfolio and blog built with React and Firebase, with a CRUD admin panel.

Saarthi: Education platform using YouTube API, MERN, Firebase, WebRTC and Socket.IO.