Feasibility assessment

Researched April 2026 · updated July 2026 with how the deferred decisions resolved.

The verdicts here aged well: CSV-first became the build plan, the "LLM never generates the numbers" sketch became the architecture's core principle, and the RIA analysis remains the reference for the day Felix charges. Where a decision this doc deferred has since been made, the resolution is recorded in place.

---

1. Technical Feasibility

Data Ingestion: Three Options

The original assessment evaluated three options and deliberately deferred the choice. It has since resolved: CSV import first (build phase 1), then SnapTrade as the US sync rail, with the CDSL/NSDL CAS statement as the no-login backbone for India — a jurisdiction this assessment didn't originally cover. Plaid remains deferred. The evaluation is kept because the tradeoffs still govern future rail decisions.

Option 1: CSV/PDF Upload (simplest — the phase-1 choice)

Users export holdings from their brokerage and upload CSV or PDF. App parses them.

• Cost: $0 in third-party fees
• Data quality: Best of all three options. Brokerage CSVs include lot-level cost basis, account types, and full holdings accurately.
• Coverage: Works for every brokerage — everyone can export a CSV.
• Build effort: Low. Write parsers for the top 5 brokerage CSV formats (Fidelity, Schwab, Vanguard, E*Trade, IBKR). LLM-assisted extraction for PDFs/screenshots.
• Tradeoff: No real-time sync. The analysis is a point-in-time snapshot. Users re-upload for updated analysis.
• Best for: Initial build, proving the analysis value prop, testing with real user data.

Option 2: Direct Broker APIs (selective real-time)

Connect directly to brokerages that offer APIs, CSV for the rest.

• Cost: $0 in aggregator fees, but engineering time per integration
• Coverage: Schwab + E*Trade covers the two biggest equity comp platforms (where RSUs live). That's meaningful for the beachhead. Fidelity/Vanguard gap requires CSV fallback.
• Build effort: Medium. Each broker API is a custom integration. Approval processes take weeks-months.
• Tradeoff: Partial real-time. Good for the equity comp accounts that change most often. CSV fills the gaps.
• Best for: When you want real-time for the highest-value accounts (equity comp) without paying aggregator costs.

Option 3: Aggregator (Plaid, MX, SnapTrade, etc.)

Single API to connect to ~12,000 financial institutions.

• Cost: $0.50-$3/connection/month × 5+ accounts/user = $2.50-$15+/user/month for Plaid-class pricing. At scale this is 10-40% of revenue at MVP pricing — which is what makes SnapTrade's model attractive.
• Data quality: Generally the weakest of the three for investment-specific data. Position-level cost basis is partial, lot-level is unreliable. 401k coverage is spotty (many recordkeepers have limited connectivity). No RSU vesting schedules. SnapTrade's brokerage-native focus mitigates this for positions, orders, transactions, and balances — exactly the data Felix needs.
• Coverage: Broadest — connects to thousands of institutions automatically.
• Build effort: Low-medium. Single integration, but OAuth flows, error handling, and connection maintenance add complexity.
• Tradeoff: Broadest reach; quality and cost vary sharply by vendor.

What no option solves (always manual):

• RSU vesting schedules — no aggregator or broker API exposes this. Manual entry or equity comp platform integration (Carta, Schwab Stock Plan, E*Trade Equity Edge — none have open third-party APIs).
• Some 401k plans — even aggregators struggle with certain recordkeepers.

How it resolved

Start with CSV — free, best data quality, proves the value prop — then layer in SnapTrade for US sync. The one reframe the original analysis missed: daily-fresh value doesn't require daily account sync. Holdings change slowly; value changes daily via prices. The architecture re-marks a slowly-synced holdings set to each night's EOD prices, so even the SnapTrade rail is pulled sparingly, and "real-time monitoring" stops being the cost driver this assessment worried about. For India, order-capable broker APIs demand daily logins that buy execution Felix doesn't use, so the no-login CAS statement is the backbone instead. See data connectivity & the daily-refresh model.

---

Core Algorithms

Portfolio Analysis Against Benchmarks

• Complexity: Low-Medium. Straightforward.
• Approach: Portfolio weights vs. benchmark allocations. Factor exposure (Fama-French). Risk metrics (Sharpe, Sortino, max drawdown).
• Libraries: Python quantlib, pyportfolioopt, empyrical, ffn. Kenneth French data library (free).
• Key challenge: Security master mapping (VTSAX and VTI are the same thing).

Tax-Loss Harvesting Identification

• Complexity: High.
• Requirements: Lot-level cost basis, current prices, wash sale rule logic (30-day window, across ALL accounts including IRAs), substantially identical security rules.
• Libraries: No robust open-source TLH library. Wash sale implementation is custom.
• Assessment: Algorithm isn't the hard part — accurate lot-level data is.

Asset Location Optimization

• Complexity: Medium-High.
• Approach: Place tax-inefficient assets (bonds, REITs) in tax-advantaged accounts; tax-efficient (index, growth) in taxable. Linear optimization problem. Academic literature is deep (Daryanani 2004, Sosner et al. 2019).
• Libraries: cvxpy for optimization, pyportfolioopt. No off-the-shelf asset location optimizer exists.

Rebalancing Recommendations

• Complexity: Medium (single account), High (cross-account).
• Nuances: Tax cost of rebalancing in taxable. Rebalancing bands. Cash flow rebalancing. Minimum trade sizes.
• Assessment: Single-account is trivial. Cross-account household rebalancing with tax awareness is genuinely hard but solved by robos. The math is known.

Backdoor Roth Guidance

• Complexity: Medium (rule-based).
• Approach: Decision tree based on income, existing IRA balances, filing status. Pro-rata rule is the key complexity.
• Risk: Getting this wrong has real tax consequences. Highest compliance/liability risk.

RSU/Equity Comp Planning

• Complexity: High.
• Requirements: Vesting schedule (manual entry), stock price projections, tax bracket projections, concentration risk thresholds.
• Assessment: High value but data entry burden is significant. V1: basic concentration risk alerting. V2: full vest-by-vest planning. Still deferred in the current build.

Retirement Projections

• Complexity: Medium.
• Approach: Monte Carlo simulation. Well-understood methodology. Open-source calculators exist.
• Key challenge: Assumptions matter enormously. Need clear disclosure and sensitivity analysis.
• Originally deferred past MVP; now in the current build's deterministic fact catalog, alongside what-if analysis.

Overall: All algorithms are feasible. Hardest problems are (1) lot-level data quality, (2) cross-account optimization, (3) RSU data ingestion. None are unsolvable.

---

AI/LLM Layer

Where AI adds genuine value:

• Natural language Q&A about your portfolio ("Why is my Roth IRA underperforming?")
• Personalized explanations of recommendations
• Scenario analysis in natural language
• Conversational onboarding/intake (better than a 20-field form)

Where AI does NOT add value (use traditional algorithms):

• Portfolio optimization, rebalancing math, tax calculations, Monte Carlo — must be deterministic. Never let an LLM do math with someone's money.

Architecture: Traditional algorithms produce recommendations. LLM is a presentation/interaction layer — explains, contextualizes, answers questions. LLM never generates the numbers.

This sketch became the built system's core principle — deterministic facts, generated words — hardened well beyond a "presentation layer": the LLM receives typed fact packs, emits schema-validated structured actions that cite fact IDs, and validators enforce that action quantities match the deterministic plans they cite.

Risks:

• Hallucination: Could fabricate tax rules. Mitigation: RAG over curated financial content, citations.
• Liability: If AI says "do X" and user loses money, who's responsible?
• Regulatory scrutiny: SEC attention to AI-driven financial advice centers on enforcement of existing fiduciary obligations and "AI-washing," not new rules — the proposed predictive-data-analytics rule was withdrawn in June 2025.
• Confidence calibration: Must express appropriate uncertainty.

---

2. Regulatory Feasibility

Is Advisory-Only Still Regulated?

Short answer: Almost certainly yes, if you provide personalized recommendations for compensation.

Investment Advisers Act of 1940: An "investment adviser" is anyone who (1) provides advice about securities, (2) as a regular business, (3) for compensation. A subscription for personalized portfolio recommendations meets all three prongs.

Not touching money does not exempt you. Custody and execution are separate concerns. Personalized advice about securities is the trigger.

The education vs. advice line:

• General education ("here's how TLH works") = NOT regulated
• Personalized recommendation ("sell THIS holding in YOUR account") = IS investment advice
• This app clearly crosses into personalized advice.

Felix's current posture

The friends-and-family deployment charges no fees, which removes the "for compensation" prong, and the system holds no execution scope at any broker by construction — advisory-only is a regulatory firewall, not just a feature. The analysis below becomes live again the day Felix charges; it is the reference for that decision, not a decision already taken.

How Competitors Handle This

Mezzi's 2026 registration is the most relevant data point: advisory-only AI startups in this space do end up registering once they charge.

Regulatory Paths

1. Register as an RIA: Most legitimate path. SEC registration for internet adviser is feasible. Compliance cost: $50K-$150K/year. Annual ADV filing.
2. Partner with existing RIA: White-label under a registered entity. Less burden but adds dependency.
3. Stay on "education" side: Provide analysis/data, frame as informational. Narrow path — legal counsel must bless every feature.
4. Disclaimer approach: "Not financial advice" disclaimers have limited legal protection if the substance is personalized advice.

Fiduciary standard: RIAs owe fiduciary duty. This actually helps positioning but creates legal exposure.

Verdict: RIA registration very likely required at monetization. Not a showstopper — known cost. Budget $50-150K/year. Get a securities attorney involved before charging.

India adds its own regime (SEBI's investment-adviser regulations) — unassessed so far, and a prerequisite before monetizing Indian users.

---

3. Resource Feasibility

This section models a funded startup build; it's kept as the reference for what a commercial build costs. The actual track is a solo, friends-and-family build — no team, no aggregator bill, roughly $85–95/month in infrastructure and AI at current defaults per the cost model.

Team Requirements (Minimum for a commercial MVP)

• 1 Full-stack engineer (app + backend)
• 1 Financial data/backend engineer (aggregator integration, analytics engine, algorithms)
• 1 Product/design (UX for complex financial data)
• 1 Compliance/legal (part-time/consultant)
• 0.5 Financial domain expert (consultant — validate algorithms, edge cases)

Minimum viable team: 3 FT + 2 PT consultants

Timeline Estimate

Note: RIA registration takes 1-4 months. File early.

Cost Estimate (commercial MVP, 12-month runway)

If founders are the engineering team: $100K-$300K for the first year.

---

4. Dependency Risks

Aggregator Risk — Low

• Not a dependency for the initial build (CSV first), and the chosen rail (SnapTrade) prices at ~$1.50/connected user/month with 5 free connections — nowhere near the 10-40%-of-revenue exposure that Plaid-class pricing would create at MVP price points.
• The nightly re-pricing model further reduces exposure: sync is periodic, not the daily heartbeat.
• Residual risk: vendor concentration on SnapTrade for US sync; CSV remains the universal fallback.

Platform Risk (App Stores) — Low

• No restrictions on advisory-only financial apps to date. Apple's 30% cut doesn't apply to web-based subscriptions.

Regulatory Risk — Medium

• SEC attention to AI-driven advice continues under existing fiduciary rules. Registering as an RIA at monetization provides a defensible position. India's SEBI regime is unassessed.

Data Accuracy Risk — Medium

• With CSV upload and brokerage-native SnapTrade data, accuracy is high — brokerage records are the source of truth.
• Mitigation: show data provenance and encourage verification against brokerage statements; the "based on" evidence chips make every computed number traceable.

---

5. MVP Scope

How this resolved

Neither option below was built as scoped. The build went straight to the full advisory loop — nightly refresh, deterministic fact catalog, daily AI briefing with structured actions, a verification loop, news, and chat. The X-ray analytics live on inside the deterministic analytics engine; the product shape is in Surfaces. The options are kept because the reasoning — especially the retention analysis — is what drove that design.

Core Hypothesis to Test

"High-earning young professionals are making suboptimal financial decisions because they don't have a trusted, accessible way to know what to improve — and they would pay for a tool that tells them."

MVP Option A: Scorecard + Recommendations

Connect accounts → Financial Health Scorecard → Top 3 Action Items → Detailed Guidance

IN V1 (must-have):

1. Account aggregation — connect brokerage + retirement accounts
2. Portfolio analysis — allocation breakdown, expense ratio analysis, overlap detection, concentration risk
3. Financial health scorecard — A-F or 0-100 across diversification, tax efficiency, fees, risk alignment
4. Top 3 recommendations — prioritized, specific, actionable, with estimated dollar impact
5. Basic tax optimization awareness — flag TLH candidates at position level
6. LLM-powered explanations — "Why is this recommended?"

Tradeoffs: More actionable but hits the advice-to-action gap harder. Closer to "personalized advice" (regulatory implications). Wider feature surface to build.

MVP Option B: Portfolio X-Ray + Analysis + TLH Flagging

Connect accounts → Cross-account X-ray → Deep analysis → TLH opportunity flagging

A tighter wedge focused on insight over instruction:

1. Portfolio X-ray — connect all accounts, unified view of what you actually hold across every institution. Allocation breakdown, fund overlap detection ("you own the same stocks in 3 funds across 2 accounts"), fee drag quantification, concentration risk, asset location assessment ("you have bonds in taxable — that's costing you").
2. Portfolio analysis — how your actual portfolio compares to where it should be. Risk exposure vs. stated tolerance, diversification gaps, benchmark comparison, expense drag in dollars/year.
3. TLH flagging — "you have ~$X in harvestable losses right now" at the position level. Identify the opportunity, explain the mechanic, suggest substitute securities. Not lot-level instructions — opportunity awareness with dollar estimates.
4. LLM-powered explanations — natural language Q&A about your portfolio. "Why do I have so much overlap?" "What's wrong with bonds in my taxable account?"

Why this scoping might be stronger:

• X-ray is the hook — instant value on first connect. Answers "what do I actually have?" before "what should I do?"
• Lighter regulatory surface — showing analysis of what you have is closer to "education" than "sell this." Still likely needs RIA, but smaller recommendation surface.
• TLH flagging adds concrete dollar value immediately without requiring lot-level data accuracy
• Sidesteps the advice-to-action gap — the product is insight and awareness, not a to-do list
• Cross-account view is the core differentiator vs. Empower (which does single-account well) and vs. Origin (which is shallow on investment analysis)

Risks with this scoping:

• Close to what Empower's free dashboard does (minus TLH). Differentiation must come from depth, explanation quality, and cross-account perspective
• If it's only an X-ray, retention could be weak — you look once and don't come back. Needs ongoing value (proactive alerts, periodic re-analysis, market-triggered TLH notifications)
• Mezzi is building something similar

Tradeoffs: Narrower, faster to build, cleaner regulatory story. But less actionable — may test "do they want insight?" without fully testing "will they pay for ongoing guidance?"

What both options deferred:

• RSU/equity comp planning (high data entry burden) — still deferred
• Crypto aggregation — still deferred
• Detailed retirement projections — now in the build
• Backdoor Roth step-by-step guidance — still deferred
• Lot-level tax-loss harvesting — the fact catalog's TLH scanner works from imported lots
• Cross-account rebalancing instructions — now in the build as structured actions

---

6. The Advice-to-Action Gap

The Problem

The target user feels overwhelmed. They know they should optimize but don't because:

1. They don't know what to do (knowledge gap)
2. They don't know if what they're doing is right (confidence gap)
3. Even if told what to do, execution is tedious (action gap)

The app solves (1) and (2). But (3) remains.

Will They Actually Act?

Honest assessment: Many won't — at least not immediately.

• Betterment/Wealthfront grew specifically because they eliminated the action gap
• Empower's free dashboard (advisory-like) converted ~2% to paid advisory+execution. The 98% who didn't convert mostly didn't act on recommendations.
• "Advice only" tiers consistently show lower engagement than "advice + execution" tiers

What Would Make Users Act? (Ranked)

1. Dollar impact quantification: "Inaction costs you $2,100/year in unnecessary taxes." Loss aversion is the strongest behavioral lever. Update this number on every login.

2. Brokerage-specific step-by-step instructions: Not "sell VXUS." Instead: "Log into Schwab → Go to Trade → Select account ending in 4829 → Enter: Sell 47 shares of VXUS at market." With screenshots for top 5 brokerages.

3. Urgency triggers: "VXUS is down 12%. If it recovers, this opportunity disappears." Time-bounded opportunities create action.

4. Progress tracking: "You've completed 2 of 5 optimizations. Annual savings so far: $3,400. Remaining: $2,800."

5. Chunking: One recommendation per week. "This week's optimization: 10 minutes, $800/year impact."

Is This a Showstopper?

No, but it shapes the product strategy materially.

• Conversion ceiling: Without execution, expect 20-40% of users to act regularly. The rest check in periodically.
• Long-term implication: Creates pull toward eventually offering execution or partnering with brokerages for "one-click" implementation. Felix resolved this the other way — advisory-only is a load-bearing regulatory firewall, enforced by holding no trade credentials at all.
• How the build attacks the gap: every recommendation is a structured action the user accepts or dismisses, and a verifier matches accepted actions against real imported transactions — the loop closes with evidence the user actually acted, and unactioned items nudge or expire instead of piling up as homework. See the action lifecycle. Whether that moves the 20-40% number is a question for real users.

---

Summary

The three decisions this doc said to make before building — and how they went

1. Regulatory path. Resolved for now by the no-fee friends-and-family track; RIA registration (and India's SEBI regime) becomes live at monetization.
2. Cost basis strategy. Decided: CSV import first, SnapTrade sync layered on — brokerage-native data over commodity aggregation, so lot-level TLH is possible from day one.
3. Execution gap strategy. Decided the other way from where the analysis leaned: advisory-only permanently, enforced by construction — the system holds no trade scope at any broker, and the verification loop substitutes for execution as the way the loop closes.