Recommender systems
Candidate generation, ranking models and cold-start strategies for feeds, catalogs and discovery surfaces — including the awkward first week when you have no signal at all.
Personalization engineering
Personalization systems, built end to end.
PersonaInn designs and ships the machinery behind personalized products — recommender engines, candidate generation, ranking, search relevance, and the evaluation that proves it actually worked. Grounded in PhD research at the University of Minnesota, and proven on platforms we run ourselves.
Recommender systems · Search & ranking · Human–computer interaction
- 2Platforms live in production
- 8+Years of research & teaching
- 0 → 1Taken from schema to deploy
- CIIR ’23Best Paper award
01 — Services
What we build
Personalization is rarely one model. It is a pipeline, a serving path and a way of knowing whether it helped. We work across all three.
Search & relevance
Hybrid retrieval, semantic matching and learning-to-rank, tuned against the queries your users actually type rather than the ones the demo used.
Evaluation & experimentation
Offline metrics chosen because they track the outcome you care about, A/B tests with guardrails, and honest reporting on what moved and what did not.
Data & ML platform
Ingestion pipelines, event and feature stores, background workers and serving infrastructure — the unglamorous layer that decides whether any of it stays up.
Full-stack product build
Schema, API, frontend and deploys. The whole product, shaped for production from the first commit rather than a prototype that has to be rebuilt.
Research advisory
Literature-grounded design, evaluation design and user studies, for teams that need the human–computer interaction side taken as seriously as the model.
02 — Method
How we work
Most personalization problems are data problems wearing a model costume. So we go looking for that first.
-
01
Discover
Audit the data you already collect and define the decision the system is really making. Often the fix is upstream of any model.
-
02
Model
Start with the simplest baseline that could work, then earn every increment against it. Offline evaluation before anything reaches a user.
-
03
Ship
Serving path, caching, background workers and monitoring. Production shape from day one, not a notebook thrown over a wall.
-
04
Measure
Experiments in front of real users, with guardrails to catch the regressions that a headline metric will happily hide.
03 — Work
Platforms we built and run
Not client logos — products we own, operate and keep in production. Every layer of these is ours: schema, API, frontend, infrastructure.
Consumer platform
ReviewInn
A review platform for the decisions that actually matter — universities, companies, places, products and professionals. Reviews are organized into groups and a trust graph, so weight follows credibility rather than volume.
- Entity taxonomy spanning five review domains
- Trust graph weighting reviewers by credibility
- Group-scoped reviews, moderation and reputation
- Django + DRF API, Next.js web, Flutter mobile
- Django
- DRF
- PostgreSQL
- Celery
- Next.js
- Flutter
Research network
ResearchAR
A social network for researchers, built around a personalized discovery feed. Your interests lead the ranking, a slice of trending keeps discovery alive, and the feed never goes empty — the failure mode that kills these products.
- Interest-led feed ranking blended with trending
- OpenAlex ingestion pipeline for papers and venues
- Points, badges and a study-participation economy
- Messaging, groups, comments and unified search
- Django
- DRF
- PostgreSQL
- Celery
- Next.js
- OpenAlex
04 — Notes
How we think about this
Interactive explainers on the parts of a recommender that are hard to picture from a formula. Every widget runs the real algorithm in your browser.
Diversity without wrecking relevance
Maximal marginal relevance with a λ you can drag — the point where six nearly identical results become six useful ones, and the point just past it where the list stops answering the question.
Reading a ranking metric: NDCG, MRR and the rest
Precision, recall, MRR, average precision and NDCG computed live from a list you can reorder and regrade — including the reorderings that move one metric and leave another completely still.
Fusing two rankers without calibrating their scores
Reciprocal rank fusion, weighted RRF and CombSUM, compared in a playground where a keyword ranker and a concept ranker disagree — and you decide how to reconcile them.
05 — Stack
What we work in
One stack, used deeply, on systems we operate ourselves — not a list of everything we have ever opened.
- Python
- Django
- DRF
- PostgreSQL
- Celery
- Redis
- Next.js
- React
- TypeScript
- Flutter
- Docker
- Cloudflare
06 — About
Led by Mahamudul Hasan
PhD candidate in Computer Science and Engineering at the University of Minnesota, Twin Cities, researching recommender systems and human–computer interaction. Before Minnesota, eight years teaching and researching in Bangladesh, most recently as a Senior Lecturer at East West University — recognized with a Best Researcher Award and a Best Paper Award at CIIR 2023.
The research and the products feed each other. What holds up under evaluation gets shipped; what real users do sends us back to the literature. That loop is the whole reason PersonaInn exists.
07 — FAQ
Common questions
What does PersonaInn actually build?
The machinery behind personalized products: recommendation and ranking systems, search relevance, the ingestion and serving infrastructure they need, and the evaluation that shows whether any of it worked. When it helps, we build the product around them too.
Do you work on existing products or start from scratch?
Both. On an existing product the first step is usually an audit of the data you already collect, because most personalization problems are data problems wearing a model costume. From scratch, we start with the simplest baseline that could work and earn every increment from there.
How do you know the personalization is working?
Offline metrics chosen because they track the outcome you care about, then experiments in front of real users with guardrails on the metrics you cannot afford to lose. A recommender that improves click-through while quietly narrowing what people see has not improved anything.
What technologies do you work in?
Python and Django with DRF on the backend, PostgreSQL as the source of truth, Celery and Redis for background work, Next.js and TypeScript on the frontend, Docker for everything, and Cloudflare at the edge. The platforms we run ourselves are built on exactly this stack.
How does an engagement start?
With a conversation about the decision you are trying to make better — what the user sees, and what you want them to get out of it. From there we scope a first piece of work small enough to be useful on its own.
08 — Contact
Tell us what you are trying to personalize.
A feed that surfaces the wrong thing, a search box people have given up on, or a product that has no idea who it is talking to — those are the conversations we like. Describe the problem and you will get a straight answer on whether we can help.