Designed and evaluated personal-memory middleware for language models, measuring state correctness, answer reliability and processing cost across nine approaches.
Based in London
Building for the real world
AI Engineer
Currently a Founding Engineer
LLM applications
Memory systems & evaluation
Pinta B2B
and ScanTrack
// Early-stage startup · London / Remote
Building Pinta Logistic B2B and ScanTrack for an early-stage startup. I own the engineering end to end, from system architecture and backend services to user interfaces and cloud deployment, working remotely from London.
LLM applications
and personal memory
// Independent work alongside my startup role
Alongside my work as a Founding Engineer, I independently build LLM applications and investigate how language models store, retrieve and use personal information. My focus is agent memory, retrieval and evaluation: testing what a system remembers, whether its answers are reliable, and what processing costs it incurs.
Explore the research ↗I build AI systems.
Then test their limits.
AI Engineer in London. Currently a Founding Engineer at an early-stage startup, with independent research in memory, retrieval and LLM evaluation.
Explore my careerFrom idea to application
Voice assistants and LLM applications, backed by Python, APIs and cloud infrastructure.
See the work ↗Memory that can be tested
Research into personal memory, measuring state correctness, answer reliability and cost separately.
Explore the research ↗Delivered business systems
Pinta Logistic B2B platform and ScanTrack: delivered services for logistics, warehouse operations and ERP integration.
MSc Big Data Science, QMUL · Distinction · 2024
B.Tech Information Technology with Data Analytics · First Class with Distinction
From models to working systems.
AI engineering and research, built on Python, backend and cloud experience.
Hover or focus a topic to select its connections. Tap on touchscreens. Choose All connections to reset.
Skills by area
Python is the central foundation connecting these areas.
AI & research
Prompt Engineering, LLM applications, Agent memory, Retrieval, LLM evaluation, SQLite
Voice applications
Pipecat, Deepgram, WebRTC, Claude API
Backend engineering
Python, JavaScript, React, Flask, FastAPI, MongoDB, PostgreSQL, Redis, Celery, Gunicorn, HTMX, 1C ERP, Hostinger, Scrapy, Boto3, Google OAuth, Toast POS, Bash, C++, SEO
Cloud & infrastructure
Docker, Kubernetes, Nginx, AWS, Cloudflare, Selectel
Data systems
Pandas, Spark, Scala, Big Data, Data Analysis
Machine learning
OpenCV, scikit-learn, Machine Learning, Deep Learning, CNN
A real-time voice booking prototype combining conversational AI, validated booking tools, persistent storage and a staff dashboard.
A conversational restaurant allergen assistant for dish-specific and multi-allergen queries, with explicit uncertainty and cross-contact messaging.
A beverage stocktaking project connecting bottle quantities, inventory workflows and review.
Full-ownership B2B platform connecting factories, warehouses, and retail buyers. Sole engineer across the entire stack - four-tier role system, Flask/HTMX frontend, FastAPI microservices, Celery + Redis async infra, and Selectel cloud deployment.
Role-based fleet and warehouse management for Pinta Logistics across 11 Russian branches. Drivers scan QR codes, barcodes, and DataMatrix labels - product details load instantly from branch-specific 1C ERP APIs. Backend split into two Gunicorn services with OpenCV image preprocessing.
Predictive maintenance platform for the Hilti SCC Hackathon - a 4-5 month international competition. HTOT collects tool usage data and runs it through a pre-trained ML model to predict service requirements before failure, reducing construction site downtime. Finished First Runner Up representing India.
Real-time ASL recognition using CNN-based image segmentation, comparing ResNet50, VGG16, and MobileNet on live webcam feed. VGG16 achieved 0.99 mAP at 60ms latency. Covers ASL alphabets and salutations - published in IJRESM Vol. 5 No. 6 (2022), co-authored at MIT-ADTU, Pune.
Pixel-perfect recreation of Angus Steakhouse's designer PDF menus as a multilingual web experience - English, French, Italian, Spanish, Chinese, and Japanese with persistent language memory. Built as an internal pitch to Noble Restaurant Group for transitioning from printed PDFs to a live digital system.
A document ingestion and retrieval prototype with an evaluation workflow for measuring supporting-evidence retrieval.
// Research and publications
Preprint · Zenodo · November 2026
Validated Personal Memory for Language Models: A Bounded Study of State Correctness, Answer Reliability, and Processing Cost
Shounak Pattewale · Published · Version 1.0
A bounded study of structured personal-memory middleware, examining state correctness, answer reliability and processing cost separately. Published preprint; not peer-reviewed.
Nine methods across three model configurations produced 1,620 answer outputs. All 486 saved state-transition traces matched their references. Separate public-benchmark analyses evaluate models without the middleware.
DOI: 10.5281/zenodo.24210707 · Creative Commons Attribution 4.0
Ongoing exploratory research
Neural Memory and Adaptive Storage
Investigating whether neural networks can store editable facts, support updates and forgetting, and preserve unrelated information. I developed prototypes, training and evaluation workflows, and comparisons with exact-record storage.
NMD002 recorded 126 completed backend runs. None of the tested weight-based backends passed the full evaluation screen; a query adapter passed its synthetic interface screen. These findings support retaining exact records as authoritative storage while exploring neural interfaces.
This earlier work informed the subsequent SQLite-backed personal-memory study. The mechanisms are distinct: no validated replacement for conventional storage or solution to catastrophic forgetting is claimed.
Journal publication · IJRESM · 2022
Comparative Study of CNN-Based Sign Language Translation Models
Co-authored sign-language research published in IJRESM, Volume 5, Issue 6 (2022), during my undergraduate studies and before my MSc. This journal publication is separate from the 2026 personal-memory preprint.
Research timeline
14 November 2026 - Published the personal-memory research preprint on Zenodo.
November 2026 - Prepared the bounded structured-memory study, building on questions raised by the earlier neural-memory experiments.
Earlier experiments - Explored neural-memory storage, updates and forgetting before the structured personal-memory study. Exact start date is not established; this work remains exploratory.
2022 - Published CNN-based sign-language research in IJRESM, before the MSc.
A useful system has to hold up beyond the demo. In production, that means dependable services and maintainable code. In research, it means testing assumptions, examining failures and being clear about what the evidence supports.
- London · Late hours · Coding · Strong Coffee
Report findings, failures and limitations.
Professional experience, education and earlier technical internships.
Own full-stack client delivery from discovery and technical planning through architecture, implementation, deployment and support. Build Python/FastAPI services, PostgreSQL-backed systems and React/TypeScript interfaces; manage Linux-hosted Docker infrastructure. Build and evaluate AI-enabled features and retrieval workflows with attention to reliability, latency and cost.
Graduated with Distinction. Studied large-scale data processing, machine learning and statistical modelling.
Developed Flask services and integration pipelines connecting CloudKitchens/Hubster with Toast, Lightspeed and other external platforms. Built and debugged REST APIs, webhooks, authentication and data mappings; investigated production integration failures and collaborated on fixes. Worked with databases, testing and Git-based delivery.
First Class with Distinction. GPA 8.17/10; third rank.
Developed backend functionality and API/database logic for an early-stage product, collaborating in a fast-moving startup environment.
Built and maintained backend API functionality with Node.js and MongoDB, collaborating through Git-based development workflows.
Worked on media morphing detection using Python, computer vision and classification models for manipulated-content analysis.
Get in touch
Get in touch about AI engineering roles, client projects or research. My focus is LLM applications, memory systems, retrieval and evaluation, backed by Python and cloud engineering experience.



