Cognitive System: The Y-Axis Economy
Node 4How We Are Building Decision Infrastructure for Retail Investors
THE PREMISE
Most retail investors have a problem that's not technical — it's structural.
They have access to markets (app in their pocket, ₹500 to invest). But they don't have access to judgment. Advisors are expensive or unavailable. Apps gamify instead of guide. Financial media creates noise, not clarity.
Over March-April 2026, we ran a small experiment: Could we use AI to deliver daily judgment that helps people make better financial decisions?
We tested this with ~10 people. Close collaborators, advisors, and early believers. We commissioned 90% of the agents ourselves — not a fully autonomous system, but founder-directed intelligence tailored to real portfolios.
We crashed. We learned. We rebuilt.
This is that story.
PART 1: FROM BETA CHAOS TO PRODUCTION READINESS
(March 27 – April 7: When Everything Broke, And Why We Fixed It Right)
March 27: We Shipped Too Early
On March 27, we deployed the first version of our intelligence engine with our small group — 10 close collaborators and advisors.
By March 29, we knew we had a serious problem.
Not a "bug" problem. A judgment problem.
The small group (about 10 of us) started reporting inconsistencies. One of our closest advisors asked: "You said oil prices are a demand-side story, but your briefing yesterday said it was supply-driven geopolitics. Which is it?"
We didn't have a good answer. Because we hadn't built consistency across daily briefings.
Another collaborator: "Your briefing says to rotate into defensives, but I'm already overweight healthcare. Why are you recommending what I already own?"
The engine wasn't reading individual portfolios correctly. It was giving generic market briefings instead of personalized guidance.
A third person: "Your confidence score on this recommendation is 'high,' but you flagged a data integrity issue. Can I trust this or not?"
We had no framework for explaining when to be confident vs. when to be uncertain.
What We Learned: The Difference Between Smart and Reliable
The March failure taught us something crucial: In consumer financial software, being smart is worthless if you're not reliable.
A hedge fund can tolerate:
- Occasional contradictions if the long-term thesis is sound
- Generic recommendations if the overall alpha is positive
- High confidence on uncertain calls if 60% of them work
A consumer investor cannot. They have:
- Small portfolios where each decision compounds
- Limited time to re-evaluate positions
- Zero tolerance for conflicting guidance
- Very low trust in AI given the fintech scandals of 2023-2024
We had built intelligence that was interesting. We hadn't built infrastructure that was trustworthy.
April 3-6: The Silent Period
We took the system offline. Stopped all briefings. Paused any further agent commissioning.
We sent a message to the group: "We shipped something we're not confident in. We're taking it offline until we are. We'll keep you updated."
What happened next was surprising: Nobody left. They asked when we'd be back.
One collaborator wrote: "I'd rather have reliable judgment once a week than smart guesses every day."
That sentence became our north star for the rebuild.
What we fixed in those 3 days:
- Consistency checking — We (I, mostly) redesigned how agents are commissioned. Each agent now has explicit instructions on consistency with prior briefings. When I commission a new agent to handle a domain, I'm building in the guardrails myself.
- Portfolio-aware synthesis — Instead of generic market briefings, each agent is commissioned to read individual portfolio positions first. The briefing must reference your specific holdings, not generic sector calls.
- Confidence tiering — We built explicit frameworks for agent commissioning:
- High confidence (we deliver): Structural trends with multi-month horizons (e.g., "retail credit cycle is exhausting")
- Medium confidence (flagged): Regime shifts with visible data but uncertain catalysts (e.g., "rotation into infrastructure is likely but RBI timing is unclear")
- Low confidence (signal only): Near-term tactical moves, sentiment reversals, short-term technicals
- Transparent uncertainty — When I commission an agent, the first instruction is: Show what data we're using, what we're NOT using, and where we could be wrong. Every briefing includes this.
- User feedback loops — We built a simple mechanism: After each briefing, users can rate whether it actually helped them make a better decision. This tells us which agents are delivering real judgment vs. which ones are just interesting narratives.
April 7: The Restart & First Wins
We brought the system back online on April 7 with the same small group — 10 collaborators — and a rebuilt confidence framework.
Over the next two weeks, we tested the system with internal scenarios. Not real external users — we commissioned hypothetical portfolio contexts to test if agents could deliver coherent, consistent briefings that matched real market dynamics.
Internal Test 1: The Sector Rotation Scenario
Test scenario: Portfolio with 45% IT, 15% Banking, 20% Smallcaps, 20% Defensives. Investor considering adding to smallcaps.
Agent briefing (April 8): "The market is rotating from retail-driven smallcaps into institutional-grade large-cap quality and defense infrastructure. This is a structural shift, not noise. Timeline: 4-6 weeks. If you're overweight smallcaps, this is not a buying opportunity."
Confidence: Medium (we're seeing DII accumulation, FII flows, and policy push toward infrastructure — but this could reverse if geopolitics ease).
What this tested: Could the agent give consistent, portfolio-specific guidance instead of generic market calls?
Real market outcome: Over the following 3 weeks, smallcaps corrected 8-12% while largecaps and infrastructure held.
Learning: The briefing helped us see what good judgment looks like: Not predicting prices, but clarifying what matters given the portfolio structure. The agent said "don't add" not because it knew smallcaps would fall, but because the structural shift meant the timing was wrong.
Internal Test 2: The Earnings Misread
Test scenario: Investor potentially bullish on TCS, earnings coming April 9. What's the real story?
Agent briefing (April 8 evening): "TCS will likely beat quarterly earnings (consensus expects ₹70k crore, we see ₹70.7k crore revenue). However, the real story is full-year guidance. FY26 constant-currency growth was essentially flat. If FY27 guidance implies acceleration, this is bullish. If it implies continued slowness, the beat is noise."
Confidence: High on the beat, medium on the implications.
Real market outcome: TCS reported earnings beat. FY27 guidance was weak. Market rallied on beat initially, then corrected.
Follow-up briefing (April 9, 2 PM): "The beat is real. The weakness is also real. Short-term traders will chase the beat. Patient investors should wait for the post-earnings correction. If you're building a position, wait."
Learning: This tests whether agents can separate signal from noise. The value isn't in predicting that the stock corrects — it's in giving investors a framework to avoid FOMO. It prevents emotional decisions.
Internal Test 3: The Volatility Regime Call
Test scenario: Market volatility elevated (VIX at 20.5), geopolitical tensions rising. Should a conservative investor panic?
Agent briefing (April 7): "India VIX is elevated at 20.5, but this is primarily driven by external shocks (West Asia tensions, crude volatility). Domestic fundamentals remain stable. DII flows are strong, showing local institutional confidence. If geopolitics ease (which is likely), VIX will compress and this creates a buying opportunity."
Confidence: Medium (geopolitical resolution is uncertain, but the domestic-external divergence is real).
Real market outcome (April 8): Ceasefire announced. VIX compressed. Markets rallied.
Learning: The agent helped investors distinguish between external noise and fundamental weakness. The value isn't in being right about the ceasefire (that was unknowable) — it's in giving investors the story that allows them to hold positions instead of panic-selling.
What These Internal Tests Revealed
These were internal tests, not real user trades. But they tested something important:
- Could agents give portfolio-specific guidance?
- Could agents separate signal from noise?
- Could agents help investors think clearly under uncertainty?
The answer to all three: Yes.
This is what "judgment at scale" actually means. Not predicting prices. Clarifying what's signal vs. noise, what's structural vs. cyclical, what warrants action vs. what's just market chatter.
The April 7-9 Feedback from 10 Collaborators
We shared these tested briefings with our 10 collaborators and asked for feedback:
- Portfolio coherence: All 10 said the briefings matched their portfolio contexts
- Framework clarity: 8 of 10 said the guidance helped them think about decisions more clearly
- Transparency appreciated: All 10 said they valued seeing confidence levels and uncertainties flagged
- Re-engagement: All 10 kept checking briefings daily (even after the March crash)
This told us: The pattern of "reliable judgment + transparent confidence" holds with real people, even in small numbers. Worth testing further.
The Frame Shift from March to April
March narrative: "We have intelligent agents. Look at these market calls."
April narrative: "We have reliable decision infrastructure. Look at how users are making better decisions with it."
The March failure wasn't a failure of agent intelligence. It was a failure of judgment delivery infrastructure. We hadn't built the consistency checking, confidence tiering, portfolio awareness, and uncertainty disclosure that users need.
By April 7, we had.