I built an AI portfolio tool that never lets the AI do the math (Asterisk)
Hey IH — final-year Computer Science (Cybersecurity) student in Malaysia here. I built Asterisk on a 2011 MacBook Pro, mostly on free-tier infrastructure, because I wanted institutional-style portfolio risk analysis to actually be accessible to retail investors, not locked behind a Bloomberg terminal.
What it does: you build a portfolio, and Asterisk runs real quant models on it — Value at Risk (parametric, historical, and fat-tailed Monte Carlo), portfolio optimization, and factor model analysis — then an AI layer turns the output into a plain-language weekly brief.
The one rule I built the whole thing around: the AI narrates, it never computes. Every number comes from deterministic math, not an LLM guessing. The AI's only job is explaining what the numbers mean in plain English.
Stack: FastAPI/Python backend, Next.js/TypeScript frontend, Supabase Postgres, deployed on Render + Vercel.
It's live at asteriskai.vercel.app — genuinely early, no users yet, looking for people willing to try it with a real (or paper) portfolio and tell me where it breaks or where the numbers don't make sense. Brutal feedback welcome — that's honestly what I'm here for more than upvotes.
Hey, saw your Asterisk post and really liked the “AI narrates, never computes” principle.
I did a quick AI visibility check and asked ChatGPT: “best AI portfolio risk analysis tools in 2026.” It came back with BlackRock Aladdin, Bloomberg PORT, MSCI BarraOne, FactSet, SimCorp Axioma, etc. Asterisk didn’t show up, which isn’t too surprising given those are huge institutional platforms with massive marketing budgets.
I think the more interesting angle for Asterisk is actually the narrower, retail side — queries like “free portfolio risk analysis tool for retail investors” or “VaR calculator for retail investors.”
That’s much closer to what you’re actually positioning against: making this kind of analysis accessible to individual investors instead of locking it behind Bloomberg-style platforms.
My guess is those searches would surface a pretty different, and potentially much less crowded, competitor set. I saw something similar recently when I checked a niche invoicing tool. The broad category looked crowded, but there was a pretty clear gap once the search was narrowed down to the actual niche.
Happy to run that check and see where the retail-focused tools are getting their visibility from, if that’s useful.
I really like the separation between using AI for interpretation and using deterministic systems for calculations. AI can be extremely useful for explaining data, finding patterns, or helping users explore information, but financial calculations are exactly the kind of thing where you want predictable results.
This architecture also seems easier to trust because the model isn’t responsible for inventing or calculating the numbers. It can focus on the parts it’s actually good at while the mathematical layer remains verifiable. I think we’ll probably see more AI products moving toward this hybrid approach instead of asking the model to handle every part of the workflow.
This is a good principle; the same thing applies to any finance tool. I've run into a version of it building mine (personal finance tracker, way less fancy math, but same idea): the second you let something guess at the actual numbers instead of just explaining them, you lose the one thing that makes a finance tool trustworthy at all.
How are you deciding how much of the quant stuff to actually surface vs keep under the hood? I'd guess most retail users' eyes glaze over at "fat-tailed Monte Carlo" even if the number underneath is solid. I'm curious how you're handling that.
"The AI narrates, it never computes" is actually a smart design principle and more rigorous than most AI-powered finance tools out there. The separation of concerns is the right call — LLMs hallucinating portfolio numbers would be genuinely dangerous, not just embarrassing.
Built on a 2011 MacBook on free-tier infrastructure is the kind of constraint that sharpens scope fast. What's the next model you'd want to add after VaR and Monte Carlo? And are you thinking freemium or something else for monetisation once you have early users on it?
The narrate-never-compute rule is the same line we drew. Our SEO scanner checks 100+ ranking factors across up to 2,000 pages. Every number comes from the crawl: response times, status codes, heading structure, schema markup. The AI writes the summary. If it hallucinated a broken link the crawler did not find, a site owner would go hunting for something that does not exist.
Where I would push harder: the boundary between narrating a number and interpreting it. "Your VaR is 4.2 percent" is narration. "This is within a comfortable range for your risk tolerance" is judgment the model is making on its own, and it reads as math to the user even though it is not.
Practical test: strip every AI sentence and check if the raw numbers plus labels are still useful alone. If yes, the narration is genuinely additive. If not, it is carrying information the quant layer should have surfaced.
You're right that "your VaR is 4.2%" and "that's within a comfortable range for your risk tolerance" are fundamentally different claims, even though both read as matter-of-fact. Your test — strip the AI sentences, check if the raw numbers are still useful alone — is a good design principle, not just a critique. I'm going to hold Asterisk's narration to that standard going forward.
On attaching the metric, horizon and date to each claim: I'd enforce that in code, not only in the prompt. Have the model return each sentence with the metric IDs it relies on, then pull every number out of the text and check it against those metrics' computed values before the brief ships. Anything unmatched gets regenerated or dropped. The failure it catches is subtle: "up 9 points week over week" is arithmetic the model did itself. Precomputing the deltas you want narrated removes most of the temptation.
This is a stronger version of what I told Stefan I'd build. Tagging each sentence with the metric IDs it relies on and validating against computed values before the brief ships — rather than just trusting the prompt to behave — closes the gap properly. Anything unmatched getting regenerated or dropped is exactly the right failure mode. Building it this way.
Excellente initiative ! Développer Asterisk, un outil d'analyse des risques de portefeuille inspiré par les standards institutionnels, tout en maintenant une séparation stricte entre les modèles quantitatifs déterministes et la génération de texte par IA, est une architecture remarquable. Utiliser de vrais calculs mathématiques pour la Value at Risk et l'optimisation tout en réservant à l'IA la traduction des résultats en langage simple permet d'éviter les risques d'hallucination. Comment gérez-vous les goulets d'étranglement liés aux performances lors de l'exécution de simulations de Monte Carlo lourdes sur une infrastructure gratuite ?
Thanks — and good question. Honestly, I keep the compute budget realistic for free-tier hosting: parametric and historical VaR run fast and cheap, and the Monte Carlo path uses a reduced simulation count with variance reduction techniques rather than a massive brute-force run. It's a deliberate trade-off — accuracy that's good enough for the use case, without needing paid infrastructure. Happy to go into the specifics if useful.
Keeping the arithmetic deterministic removes one failure mode, but the narration can still create another. The AI might preserve every number while overstating what it means, hiding a model assumption, or turning a risk estimate into something that sounds like a recommendation.
I’d make each claim in the weekly brief traceable to the exact metric behind it, including time horizon, data date and key assumption. For example: “Concentration risk increased because X moved from 18% to 27%,” with the calculation visible.
The trust test is not only “is the number correct?” It is also “could the user see why this sentence was justified and where it might stop being true?”
That's a fair and important point — deterministic math handles the "is the number right?" problem, but not the "did the sentence overstate what the number means?" problem. I like your framing a lot. Next iteration will attach the exact metric, time horizon, data date, and key assumption to every claim in the brief — basically making each sentence auditable back to its source, the way you described. Appreciate the concrete direction.
That auditable-claim layer is going to pay off. Excited to see how it lands with early users once every sentence can point back to metric + date + assumption.
The deterministic-math rule is a strong trust thesis. Have any users trusted the weekly brief enough to use it for an actual portfolio decision, rather than just checking the numbers?
Honestly, no — not yet. Zero users so far, so nobody's trusted it enough to act on a real decision. That's actually the test I'm hoping to run next: get a handful of people using it with a real or paper portfolio and see if the brief holds up under actual scrutiny, not just a first look. If you're open to trying it, I'd genuinely value your read on it given the question you just asked.
That real-decision test is the important next step. If you’re open to it, what’s the best email to reach you on?
The split between deterministic models and AI narration is a strong trust choice. How will you validate the outputs with early users—one portfolio end to end, or against a benchmark?
That's a fair and important point — deterministic math handles the "is the number right?" problem, but not the "did the sentence overstate what the number means?" problem. I like your framing a lot. Next iteration will attach the exact metric, time horizon, data date, and key assumption to every claim in the brief — basically making each sentence auditable back to its source, the way you described. Appreciate the concrete direction.