Loan Labs01
Less paperwork.
More possibility.
Lois brings document classification, policy validation, and permission-aware agent actions into mortgage workflows.
LoisAgentic mortgage workflows
Explore the engineering Product engineer · Building with AI
Ideas into products.
Curiosity into work.
I’m Ashwin. I build AI products, from the systems underneath to the experience people use. I’m interested in the whole journey: the idea, the design, and what makes someone choose it.
Engineering, design, and a little
productive obsession.
Currently building at Loan LabsPreviously Finally · Virginia Tech
The lab · Try an idea
A portfolio you can use.
Small experiments in how I think and build.
Following my curiosity
The whole product interests me.
Engineering is where I build. Product design, sales, conversion, marketing, and brand are questions I keep coming back to: why does this matter, how should it feel, and why would someone choose it?
Some of this is shipped work. Some is what I’m exploring next. I keep the distinction visible.
Selected work.
Real workflows. Considered systems.
Less paperwork.
Finally02
From transactions
to understanding.
Classify AI combines retrieval, transaction history, and bank integrations to make bookkeeping less manual.
Classify AIFrom prototype to production
Explore the engineering Selected outcomes
at Finally
50K+transactions processed daily
~80%less manual categorization
$3M+in credit underwritten
More things I’ve built 7 projects
Gurukul
LLM-enhanced adaptive CS learning platform connected to published research on RAG, guardrails, and computer-science education.
Neuralflow
Full-stack productivity application combining task management and focus workflows.
Agentic systems
Public GitHub work includes a code-review agent, note-taking agent, AI transaction-classification work, and personal command-center projects.
Hemingway
A one-stop NLP web application to summarize, analyze, and paraphrase text. Python, Transformers, Streamlit.
Code Mentor
CodeBERT and CodeT5 system that explains LeetCode solutions and recommends similar challenges from vector embeddings.
Video2MP3 Distributed System
Authenticated media conversion service using microservices, Pika, RabbitMQ, Flask, PyMongo, and GridFS.
VT Search
Web search and summarization system using an inverted index and Transformers-based summarization.
Experience.
From research to real-world products.
Loan LabsApplied AI EngineerJan 2026 – Present
- Building Lois at an early-stage mortgage technology startup for internal and pilot workflows.
- Re-architected Lois from one-off Ruby LLM calls into a LangGraph agentic system on Amazon Bedrock AgentCore for mortgage-document classification, lender-specific renaming, and policy validation.
- Built agent-facing Rails APIs, borrower email intake, and an in-product conversational interface so users could handle loan documents and initiate agent actions in LoanOS.
- Designed fail-closed authorization for Composio integrations: tenant/owner scoping, separate write/send/merge/archive permissions, reviewed tool allowlists, and execution-time checks that invalidate revoked access.
- Connected CRM and document workflows across Google Drive, Box, Salesforce, HubSpot, Pipedrive, OneDrive, and SharePoint with fine-grained access controls and blocked delete actions.
- Introduced an AI-assisted software delivery workflow linking Linear/Notion product context, technical specifications, shared company knowledge, and senior-engineer code review.
FinallyAI Product Engineer / Tech LeadFeb 2024 – Jan 2026
- Joined as Finally’s first AI Product Engineer and led a three-engineer team that took Classify AI from prototype to production; built an LLM-assisted bookkeeping workflow processing 50K+ daily transactions and reducing manual categorization by ~80%.
- Designed a retrieval-augmented transaction-classification system using LangChain, Pinecone, Elasticsearch, Redis, Celery, Django, and W&B; combined semantic transaction history, merchant enrichment, and custom charts of accounts for bookkeeping recommendations.
- Evolved Classify AI from CSV upload to automated bank-data ingestion with Plaid/Teller, OCR-supported statement processing, reconciliation checks, and QuickBooks push; helped reduce first-month close time from 4+ months to ~2 weeks.
- Architected reusable banking infrastructure for account linking, token lifecycle management, encrypted access-token storage, webhooks, transaction synchronization, statement retrieval, and provider-normalized account data.
- Built cash-based underwriting for Finally’s corporate-card product using 90-day bank data, daily-balance reconstruction, Z-score logic, weekly recalculation, audit history, notifications, and manual-override controls. Underwrote $3M+ in credit for 50+ companies in approximately three months.
UNAR LabsMachine Learning EngineerJun 2023 – Aug 2023
- Developed accessibility-focused backend and data pipelines for visually impaired users using OpenCV, PyTorch, Transformers, FastAPI, Docker, GCP, and Hugging Face.
- Architected and optimized data pipelines for preprocessing, model training, and inference.
- Deployed scalable ML solutions on GCP with Docker containerization and Hugging Face model integration.
OutreachMachine Learning Platform Engineer InternMay 2022 – Aug 2022
- Built reusable NLP inference and deployment infrastructure with PySpark, MLflow, ONNX, NVIDIA Triton, Docker, Go/Python microservices, CI/CD, and GKE.
- Developed PySpark + MLflow pipelines for text processing and served ONNX models on NVIDIA Triton.
- Built Golang/Python microservices with Docker, unit tests, and CI/CD.
Research & writing
A little deeper.
Let’s work together
Something worth
building?
I’m interested in product engineering, applied AI, and founding engineering roles. Especially with people who care about how a product is built, experienced, and brought to market.
Let’s talk