4 · App Spotlight — AI Institute

Part 4 of 6 · ~9 min read

If Infinite Research is “ask one question, get a tree of answers,” then AI Institute is something quite different: a standing team of AI analysts, each with a defined role, working to a shared schedule, on a specific domain — investment research.

Think of it as a small research firm where every analyst is an AI, the office runs 24 hours a day, and the whole operation costs roughly what one junior human analyst’s health insurance would cost — meaningfully less than the analyst themselves.

This article walks through what AI Institute is, who’s on the “team,” what they produce, and what it actually feels like to operate.

Why investment research is the perfect testbed

Investment research is interesting as a use case because it has all the properties an AI orchestration platform is built for:

It has clear roles. A real investment shop has equity analysts, credit analysts, derivatives strategists, ESG specialists, compliance officers, an investment committee chair. Each has a defined job, a defined output, and a defined audience. AI workers can be assigned roles cleanly.
It has clear cadence. Daily morning briefings, weekly themed reports, ad-hoc deep dives. Most of it is scheduled work that happens on the same rhythm regardless of who’s available.
It has a clear quality bar. A well-formatted report with sources, structured analysis, and a recommendation is unambiguously useful. A broken layout or fabricated source is unambiguously not. You can tell quality at a glance.
It rewards integration. A good morning briefing pulls together what the macro analyst saw overnight, what the equities team flagged, what the compliance officer’s monitoring system caught. The value comes from synthesis — exactly what coordinated AI agents are good at.
The marginal cost of an extra analyst is meaningful at human scale, nearly zero at AI scale. Hiring an additional human ESG analyst costs $150K/year and three months. Adding an AI ESG analyst costs zero incremental dollars (the underlying Claude / ChatGPT / Gemini subscriptions are flat-rate and cover all the analysts together) and 30 minutes of prompt design.

That last point is the unlock. AI Institute exists because once you’ve paid for the AI subscriptions, adding the 6th, 7th, 8th specialized analyst is free.

The roster

AI Institute currently runs with these defined roles. Each is a specific AI worker with a specific persona, expertise focus, output format, and schedule.

Role	Cadence	Output
首席量化师 (Chief Quant)	weekly	Multi-factor models, alpha mining, portfolio construction notes
能源行业分析师 (Energy Analyst)	daily	Oil/gas/coal/renewables/electricity sector tracking
公用事业分析师 (Utilities Analyst)	daily	Power, water, environmental utilities sector
房地产分析师 (Real Estate Analyst)	daily	Property sales, land auctions, REIT tracking
可转债分析师 (Convertible Bond Analyst)	daily + alerts	Convertible valuation, premium compression, force-redemption alerts
衍生品策略师 (Derivatives Strategist)	daily + alerts	Basis trade monitoring, options flow
ESG分析师 (ESG Analyst)	daily + weekly	Carbon policy tracking, greenwashing scans, ESG score updates
合规官 (Compliance Officer)	daily stress test	Reads the day’s outputs, flags anything that needs disclosure or escalation
主题研究员 (Theme Researcher)	weekly	Themed deep dives — current themes include AI value chain, semiconductor, low-altitude economy
社交媒体分析师 (Social Media Analyst)	daily	Sentiment from Xueqiu, Eastmoney, Weibo
技术分析师 (Technical Analyst)	daily	Chart patterns, MA systems, volume analysis
投委会主席 (Investment Committee Chair)	daily morning	Synthesizes all the above into a single prioritized briefing
日报总编 (Daily Brief Editor)	daily evening	Assembles the published version of the day’s morning briefing

That’s 13 distinct analyst roles, each scheduled to run automatically. A real investment shop with this much coverage would have 30+ humans.

A typical day

Here’s what actually happens in AI Institute on a normal weekday.

06:30 — Roles wake up

A scheduler triggers. The energy analyst, real-estate analyst, ESG analyst, social-media analyst, and technical analyst all start their morning runs in parallel. Each pulls overnight data, does its specialized analysis, and writes a short brief into a shared workspace.

Each brief takes 5–10 minutes. They run on Claude or Codex CLI agents distributed across a couple of Mac Studios.

07:30 — Specialized briefs land

Six or seven 1-page reports show up in the workspace. Each has the analyst’s name, the date, a tight summary, key data points, and any recommendations.

The compliance officer wakes up next. It reads every brief produced so far, looks for anything that’s too speculative, anything that references material non-public information, anything that needs a disclaimer. It flags issues.

08:00 — The morning huddle

The investment committee chair (also an AI) reads everything: the specialist briefs, the compliance flags, the social-media sentiment report. It writes a 2-page coordination memo:

Top 3 takeaways
One actionable recommendation
Open questions for the day

This is the document I read with my morning coffee.

09:00 — Themed reports

If it’s Monday, the theme researcher kicks off a deeper investigation on whatever weekly topic is queued: “AI infrastructure in China,” “low-altitude economy in 2026,” “convertible bond opportunities post- rate cut.” This produces a much longer report by end of day.

Throughout the day — Alerts

The convertible-bond analyst and derivatives strategist run alert-mode workflows. If a convertible bond hits a force-redemption threshold or a basis trade opens up beyond a configured spread, an alert fires.

18:00 — Daily brief

The daily-brief editor compiles the morning huddle, intraday alerts, and key follow-ups into a published version. It saves to the workspace and (optionally) emails to a configured distribution list.

Overnight — Recovery and prep

The system queues up tomorrow’s morning runs. Any failed jobs from the day get retried. The workspace gets cleaned up.

The next morning, the cycle repeats.

What this looks like to me, the operator

I open my browser at 08:00. I see:

A morning briefing from the investment committee chair (2 pages)
Six or seven specialist briefs (1 page each)
A compliance-officer note (often empty, occasionally interesting)
Yesterday’s themed report if there was one
A handful of alerts since I last looked

I read the briefing. If something looks wrong or shallow, I drag the underlying brief into the kanban and write a one-line note: “this needs a deeper take on the policy angle.” That triggers a re-run with the additional context.

If a theme catches my eye, I click “research this idea” — that spawns an Infinite Research session (the app from the previous article) on the topic. Twenty minutes later I have a 30-page deep dive.

If I want to add a new analyst role — say, a “quantitative credit analyst” — I create a workflow definition: specify the prompt, the schedule, the model, the output format. Save. Tomorrow morning at 06:30, the new analyst starts producing.

That’s the operator experience. It feels like running a small research firm without the HR department.

How much it actually costs

Honest accounting for one month of running AI Institute at this scale:

AI subscriptions — the dominant line item. I run top-tier plans on Claude, ChatGPT, and Gemini, each ~$200/month. Roughly $600/month for the three combined. These are flat-rate plans that I’d be paying for anyway as a working professional; AI Institute just leverages them harder than a typical user would.
Hardware — a couple of Mac Studios and a MacBook I already own. No new hardware was purchased for this. Marginal electricity cost is small (~$10/month).
Cloud orchestration — under $5/month on Cloudflare (Workers, D1, R2, KV, Queues — all on the cheap tier).
Third-party APIs — ~$20–40/month for various data and tooling endpoints (search, embeddings, occasional specialized model calls).

Total run-rate: roughly $650/month all-in for a 13-analyst, 24/7 research operation, dominated by AI subscriptions I’d carry regardless.

For comparison, a single Bloomberg terminal subscription is ~$2,500/month and gives you data, not analysts. An equity analyst at a buy-side firm in Singapore is $200K/year fully loaded — about $17,000/month. AI Institute runs at ~4% of that cost.

The cost-per-output here is probably 100× lower than the human equivalent for the categories of work AI is genuinely good at — which is “first-pass synthesis, pattern recognition across published sources, structured summarization.” It is not better at original contrarian thinking, primary research, or deal sourcing.

What this is not

To set expectations clearly:

Not a trading system. Nothing here places orders. The output is documents for a human to read.
Not a substitute for primary research. Everything an AI analyst knows comes from published sources. It can’t talk to management, attend conferences, or visit factories.
Not a guaranteed-correct system. AI workers occasionally miss things, occasionally over-confidently state wrong things, occasionally just produce junk. The compliance officer catches some of this. The human reading the briefing catches the rest. Nobody should run capital based purely on the output without their own judgment layer.
Not a regulated investment-advice service. It produces research for personal use. If anyone wanted to commercialize this, there would be a long compliance journey before public release.

What it is is a productivity multiplier for someone doing their own research. It saves me 4–6 hours a day of routine “stay current on the markets” work that I now have done for me before I wake up.

What I learned building it

Three things I didn’t expect:

1. Roles are more important than models

Early on I assumed “use the best model for everything.” I was wrong. The same model assigned to different roles with different prompts produces wildly different output quality. A “compliance officer” persona produces meaningfully better risk-flagging than the same model asked the same question without the role framing. Persona engineering ended up mattering more than model selection.

2. The hard part is the schedule, not the analysis

Building 13 analysts that each work well in isolation took maybe 20% of the effort. The other 80% was orchestration: making sure they run in the right order, that the synthesizer waits for the specialists, that failures are caught, that the morning briefing actually shows up at 8:00 every day. The “operating system” was the hard problem; the analysts were the easy one.

3. Quality goes up dramatically when analysts read each other’s work

The single biggest jump in usefulness came from having later-in-the- sequence analysts read what earlier ones produced. The compliance officer reading the energy analyst’s draft is dramatically more useful than the compliance officer working in isolation. Cross-pollination between AI workers is where most of the leverage is.

Where this goes

Three directions for AI Institute over the next few months:

Memory across days. Right now each day starts fresh. Adding a real persistent memory (the analyst remembers what they wrote yesterday) will make the briefings much sharper.
External data sources. Wiring in real-time market data feeds, news APIs, and SEC filings will move the operation from “smart summary of public knowledge” to “smart summary of fresh public knowledge.”
Human reviewer pipeline. A queue where I can flag a brief as “needs revision” and the AI re-generates with my note attached. This is mostly UX work; the platform supports it already.

If those land, the operation becomes meaningfully more useful — and plausibly something I’d consider opening up to a small group of beta users.

The platform that AI Institute runs on isn’t sophisticated, but it took some doing to put together. The next article is the personal story of building it — what it actually feels like to construct infrastructure with an AI partner, in 10 days.

Previous: App Spotlight — Infinite Research · Next: The Build Story

AI Institute — The Research Platform