Strategy where technology and finance meet.

Project 01 · Financial AI · Local LLM Systems

Cadence: private financial intelligence from public earnings call transcripts.

A working local-first AI tool that turns an earnings call transcript into a structured executive dashboard in about 60 seconds, fully on-device, with no cloud API call.

Premise: public companies file 4 earnings calls a year, over 2,000 S&P 500 calls annually, each 15,000–25,000 words of dense forward guidance, risk disclosures, management tone, and KPI prints. Today the options are bad: manual analyst review (2–4 hours per call, subjective, can't scale across a portfolio), expensive analyst-product subscriptions (Bloomberg, FactSet, AlphaSense), or cloud-API LLMs that send market-sensitive transcripts to a third party. Each option fails on at least one of speed, cost, privacy, or scale.

A working local-first AI dashboard: ingest a transcript (.txt or .pdf) and return a structured executive view in about 60 seconds on a CPU-only laptop, no GPU required. Output: executive summary with green / amber / red sentiment badge, KPI cards (sentiment, confidence, theme count, risk count), theme-importance bar chart, risk-severity heat chart, bull / bear case decomposition, and evidence quotes with tag pills. Stack: Streamlit + Ollama running Qwen2.5 7B-Instruct + Python PDF parser + JSON-schema constrained decoding + Plotly. The product point is the local-first tradeoff: cloud APIs answer faster (about 5s against 2–4 min) and reason more deeply, while local inference keeps the transcript on the machine, runs at zero marginal cost, holds up offline, and gives full control over prompt and schema. North-star: time-to-insight per call. Counters: summary accuracy, hallucination rate, schema-drift incidents, analyst-override rate.

Project 02 · Digital Transformation · Financial Services

Vanguard: cost-to-income transformation at a mid-cap bank.

Eighteen-month digital-operating-model transformation at a Canadian mid-cap bank, recovering $130 to 170M / yr in run-rate cost-to-serve through self-serve-by-default journey routing.

Premise: the Canadian mid-cap branch bank is being consolidated out of existence. RBC absorbed HSBC Bank Canada in March 2024; National Bank absorbed Canadian Western Bank in February 2025; Laurentian agreed in December 2025 to sell. Cost is why. Laurentian, the branch-based archetype, ran a 75.2% adjusted efficiency ratio in fiscal 2025, while EQB ran 50.9% on the same fiscal year without a retail branch network. The Big Six sit between at about 54% median, with National Bank best in class at 51.2% adjusted. They already spend on technology; the gap is an operating model that still routes 70%+ of customer journeys through branch and call-center while challengers route 80%+ through self-serve. Compressing 13 to 17 points of that gap is worth $130 to 170M / yr in run-rate cost-to-serve at this scale.

Vanguard is an 18-month digital-operating-model transformation: a cost-to-serve diagnostic by customer journey (originations, servicing, claims-of-error, lending decisions), target-state operating model design with self-serve-by-default routing, vendor and platform strategy (cloud-core vs. legacy mainframe), and an agile pilot program on the three highest-volume journeys. North-star: cost-to-income ratio movement (basis points) by quarter. Counter-metrics: NPS by journey (cannot decline by more than 2 pts), digital-channel adoption rate by segment, and operational-risk-event frequency. Wedge: replatform the originations journey first, where the ROI is clearest and the risk surface is best contained.

Project 03 · Markets · Strategy & Product

Manifold: standing up a new market for a market-maker.

A Strategy & Product market-entry engagement: take an electronic market-maker into North American natural gas, financially-settled basis first, by clearing every operational, regulatory, and infrastructure launch-blocker before the first quote.

Premise: a leading electronic market-maker wants into North American natural gas. The pricing edge is the easy part. Entry is gated by three things a trading model never touches: a market structure that trades as an exchange-and-OTC hybrid priced off Henry Hub plus a lattice of basis hubs; an operational and regulatory stack (CFTC position limits, large-trader reporting, ISDA onboarding, and, for physical, FERC and pipeline scheduling); and connectivity, curve construction, and risk systems that have to speak basis, not just outright. The pricing model is the straightforward part. The launch risk sits in the plumbing around it.

Manifold runs the entry as a sequenced Strategy & Product engagement rather than a big-bang launch: shadow the desk and map the microstructure, scope the operational, regulatory, and technical integration against a single go-live date, then ship a thin-slice pilot, one liquid hub, financially-settled and cleared only, with physical deferred behind its own gate. Each workstream resolves to a weekly go / no-go with named launch-blockers, routed through a readiness frame of three gates and three states each. North-star: time-to-first-quote on a single hub. Counter-metric: zero physical exposure in the pilot, so delivery obligations stay out of scope until they earn their own gate.

Project 04 · Commercial Due Diligence · B2B SaaS / PE

Cipher: commercial diligence on a $400M vertical-SaaS target.

Four-week commercial diligence sprint on a $400M-ARR vertical-SaaS target: TAM triangulation, cohort retention read, competitive-moat test, partner-grade IC memo.

Premise: a North American mid-market PE acquirer is evaluating a $400M-ARR vertical SaaS company serving construction-trades businesses. Three things are unverified in the seller's materials: (1) the TAM number assumes the trades-software TAM is undifferentiated, when in fact the addressable wedge is 25% of the headline figure; (2) NRR of 122% is real, but cohort-level retention is degrading at the lower end of the customer base, signaling pricing-power compression; (3) the competitive moat is described as "feature breadth" but reference calls suggest the moat is actually integration-density with a fragmented field-service tooling ecosystem.

Cipher is a four-week commercial diligence sprint: customer reference call program (40 calls, weighted by ACV cohort), bottom-up TAM triangulation against three independent data sources, win/loss analysis against the four most-cited competitors, and a pricing-power test using the seller's own discount-and-uplift data. The deliverable is a partner-grade memo with a bear/base/bull recommendation and explicit deal-modifier flags (price chip, structure change, walk-away). North-star deliverable: a recommendation the IC can vote on without further analytical work; a clear answer to "what would have to be true."

Project 05 · Fintech SaaS · Agentic AI

Copilot Ledger: agentic bookkeeping for SMBs.

Agentic SMB bookkeeping built around the licensed accountant: signed audit trail, materiality-routed approval queue, channel-led GTM through firms.

Premise: SMB bookkeeping is a $25B+ category in North America, dominated by QuickBooks Online (about 80% share) and a long tail of bank-rule-based reconciliation. The category is overdue for an agentic rewrite, though the dominant pattern (full-autonomy "AI accountant") gets the regulatory model wrong: in the US and Canada, signed financial statements require a licensed accountant of record. The product wedge keeps that accountant as the signer and makes them 5× faster on the low-judgment 80% of the work.

Copilot Ledger is an agentic AP/AR + close + reconciliation product designed around the licensed accountant. Agents take responsibility for transaction categorization, vendor matching, and invoice intake; humans approve materiality-flagged exceptions, sign the close, and respond to client questions. Every agent action emits a signed audit trail (action + supporting documents + accountant reviewer ID) so the work product is defensible to the IRS, CRA, and PCAOB-style audits. GTM is channel-led: licensed-accountant firms (Intuit ProAdvisor, CPA.com partner channel) sell into their SMB book of business. North-star: close days saved per accountant per month. Counters: material misstatement rate per 10,000 transactions (cannot exceed manual baseline), audit-trail completeness, and accountant-over­ride rate.

Project 06 · Cloud Data · Marketplaces

Lakehouse Marketplace: governed data exchange across clouds.

Cross-cloud governed data exchange: usage-based metering, clean-room privacy by default, agent-aware policy enforcement at query time.

Premise: every Fortune 1000 enterprise runs 50–500 ongoing data exchanges; median time-to-production is 14 weeks at $80K–$300K per exchange per year. Platforms like Snowflake Marketplace, AWS Data Exchange, and Databricks Marketplace solved discovery; they stop short of the four hard problems that follow: cross-cloud federation, regulated-data clean rooms, producer-side monetization, and (newly urgent) agent-aware policy enforcement at query time.

Turns data sharing into a six-minute listing: usage-based metering at the credit-consumption level, clean-room privacy by default, cross-cloud federation as the architectural default, and a query-time policy plane that lets producers grant AI agents the same access as humans without losing row/column-level controls or lineage. Producer console treats every dataset as a SKU; consumer experience is "subscribe in 30s, query in 5min." Every agent query carries a signed audit record back to the data owner: the policy layer is the moat that makes regulated-data sharing safe enough to scale. North-star: compute revenue attributable to shared listings ($M / quarter). Liquidity counters: % listings with > 5 active subscribers, time-to-first-query, clean-room workloads / quarter, and policy-violation incidents per million queries (audited monthly). Wedge: convert customers' largest in-flight bilateral data exchanges into platform-listed SKUs.

Project 07 · D2C Insurance · Post-sale Lifecycle

Lifeline: the customer account as a renewable product surface.

Post-sale customer account run as a product surface owned by Product: life-event signals turn into coverage-gap nudges and cross-sell loops, NRR > 118%.

Premise: D2C insurers spend $400–$900 fully-loaded CAC per bind, absorb heavy early lapse, and convert under 6% into a second product over the lifetime. Every life event that should change coverage (marriage, child, mortgage, salary jump, divorce, claim) is observable in the customer's life and invisible to the carrier; the customer account UI is a PDF download and a billing date. Cross-product attach sits in the marketing budget instead of the product surface, and claims, the moment of highest emotional bandwidth, run through ops as a cost center when product could treat them as a growth event.

Lifeline turns the post-sale surface into a product owned by Product, not Marketing. Three surfaces, one event spine: the customer account shows current coverage against detected life context with one-tap adjust within risk class; the CS console shows the same context with a drafted next-best-action, audit-trailed; the claim flow is instrumented as the highest-NPS moment in the product. Open-banking, payroll, and customer-declared signals enter as signed envelopes; an immutable audit ledger doubles as the experimentation substrate, regulator-grade. North-star: net revenue retention on the existing book, audited monthly against a hold-out cohort: NRR > 118% from book actions alone. Counters: coverage-gap precision > 92%, cancel-save rate > 22% (catches nudge fatigue), claim-NPS > 65.

Project 08 · Search & Ads · Agentic Commerce

Atlas: when search becomes action.

Parallel ad auction over agent-readable inventory, with cryptographic receipts and a 35% publisher revenue share, capturing the $12B / yr stranded as agents replace clicks.

Premise: search ads rest on assumptions that are breaking: a human reads the SERP, a click signals real intent, and publishers share the upside. By 2027–28, projected ≥30% of high-intent commercial sessions will be at least partially agent-mediated (Gartner / eMarketer agent-commerce forecasts); ad and publisher revenue from those sessions is currently zero.

Atlas is a parallel ad platform: human SERP (existing inventory, defended) and agent inventory (a new auction over agent-readable units). Agents return cryptographically-signed receipts for downstream conversions, so verified attribution becomes a contractual artifact. Publishers whose content informed an agent's answer collect a defined revenue share (a 35% pool, daily-granular, exposed in a publisher console). North-star: verified agent-mediated conversions / week. Counter: brand-safety incidents per million impressions, capped at 1.5. Sequencing: publish the open Agent Ad Protocol before the bidder ships; own the standards layer first, the auction second.

Project 09 · Creator Economy · Marketplaces

Loop: creator monetization that compounds.

Mid-tail creator earnings as a financial product: viewer-subscription-funded floor with ±10% next-month prediction SLA, brand-safety AI defensible to a CMO.

Premise: mid-tail creators (50K–2M followers) on short-form platforms earn $400–$2,800 / month at a coefficient of variation north of 0.7. Their churn comes from financial-planning collapse rather than low reach: 18–25% quarterly churn forces platforms to spend 2–4× LTV reacquiring replacement supply. Unpredictability is the real driver.

Loop turns creator earnings into a stable monthly floor: predicted within ±10%, funded from viewer subscriptions, with brand-safety AI that's defensible to a CMO. Treat earnings as a financial product: stable floor + variable upside (performance ads + brand sponsorships, matched at content-embedding level). Per-creator algorithmic transparency in a console committing to ±10% on next-month floor. Recommender PM craft: balance creator/viewer/advertiser metrics as one objective. North-star: median mid-tail subscription floor ($/mo). Counters: brand-safety incidents per million impressions (CMO leaves at 2.0); floor-prediction accuracy (> 92%); paired-cohort q-churn (< 12% vs 22% control).

Project 10 · Cloud FinOps · Enterprise SaaS

Aperture: workflow-shaped FinOps for enterprise infrastructure.

Multi-account enterprise cloud-cost intelligence sequenced as a workflow product (commitment engine, chargeback inverter) rather than a dashboard. Audit pressure becomes the wedge.

Premise: Fortune 500 cloud spend has crossed $400B / yr aggregate, growing 25–35% YoY. The dominant tooling pattern (cloud-native cost dashboards plus a long tail of FinOps point tools) has plateaued in value capture: enterprises that have run a FinOps program for > 18 months still report 25–40% of cloud spend as untagged, mis-allocated, or under-utilized. The next-leg unlock is workflow-shaped intelligence that converts cost insight into named owner actions with named due dates.

Aperture is a workflow-shaped FinOps product: a commitment engine that converts on-demand spend into right-sized reserved + savings-plan portfolios with named-owner accountability and chargeback rails by business unit; a chargeback inverter that lets engineering leaders see their team's cloud spend the same way Finance sees the bill; and a tagging lookahead that flags untagged resources at provision time, before the month-end audit. Audit pressure (SOC 2, ISO 27001, cloud-spend disclosure under SEC Reg S-K) becomes the wedge that converts the buying center from "cost saver" to "compliance lever." North-star: committed-vs-on-demand spend mix by quarter. Counters: tag completeness at provision, untagged spend / total spend, and time-to-fixed once a savings opportunity is named.

Project 11 · Private Equity · Fund Commitments

Keystone: commit or pass on a $2.5B mid-market buyout Fund V.

A six-week LP underwrite of a buyout re-up: track-record decomposition, public-market-equivalent (PME) benchmarking, and a value-creation attribution, delivered as a sized commit/pass to the investment committee.

Premise: a pension plan's private-equity program is offered a $75M commitment to Fund V of an incumbent North American lower-mid-market buyout manager. A re-up is the default; the diligence job is to falsify it. Three priors must be priced independently: a 56% fund-size step-up ($1.6B to $2.5B) that tends to compress relative returns and can invite style drift up-market; a realized return stack still leaning 40% on multiple expansion, a tailwind that's hard to count on in a higher-rate regime; and a largely-unrealized Fund IV whose 1.7x TVPI is still a paper mark.

Keystone underwrites the manager rather than the marketing: an independently rebuilt deal-by-deal cash-flow model (KS-PME, direct alpha, loss ratio, gross-to-net bridge), a value-creation attribution decomposing gross value into revenue, margin, multiple, and deleveraging, and a terms-and-team review (GP commitment, key-person, succession, fee/carry). North-star deliverable: a commit/pass the IC can vote on, backed by an independent track-record model rather than the GP's marked TVPI, with a portfolio-construction covenant and a side-letter ask list attached.

Project 12 · Private Equity · Co-Investments

Fulcrum: underwriting a co-investment in a $520M buyout.

An independent LBO underwrite of a fee-free, carry-free co-investment alongside a sponsor: operating case, leverage and returns model, and a sensitivity that locates where the deal breaks, as a go/no-go.

Premise: a sponsor on the LP's approved-funds list offers a $40M co-investment alongside its $260M equity check in the $520M-EV buyout of a North American specialty-distribution platform, at 10.0x EBITDA and 5.0x leverage. A no-fee, no-carry co-invest lets the LP keep the gross deal return instead of the fund's 1.75% fee and 20% carry, an edge worth roughly 7-8% of IRR (Cambridge Associates) that holds net only for a diversified, independently-underwritten co-invest book, on the sponsor's clock.

Fulcrum rebuilds the LBO straight from the data room: an operating case lifting EBITDA from $52M to about $78M over five years on 6% revenue growth plus 150 bps of margin, an equity value bridge separating EBITDA growth from multiple change and debt paydown, and a sensitivity grid across exit multiple and EBITDA CAGR. North-star deliverable: a go/no-go that clears the hurdle under a one-turn-down exit and a higher-for-longer rate path, with operational value as the largest bar in the bridge, ahead of leverage, conditioned on a rate hedge and information rights.

Project 13 · Industrial AI · Energy & Operations

Wellhead: industrial AI for upstream operations.

Physics-informed edge AI for oil and gas operations, with regulator-ready ESG signing built in; survives 30+ days disconnected and learns across operators on aggregated gradients rather than raw telemetry.

Premise: 1.2M producing wells operate globally, and unplanned downtime runs into the tens of billions a year across heavy industry. The 2018–2023 pure-ML wave failed because reservoir engineers reject outputs they can't mechanistically defend. Methane regulation has hit calendar deadlines (EPA NSPS OOOOb/c, EU Methane Reg, OGMP 2.0) that require operator-signed emissions lineage, and operators file that separately today. The disconnect between operational AI and regulator-grade ESG is the gap.

Wellhead closes it: a physics-informed failure-prediction runtime (not black-box ML), edge-deployed so it survives connectivity loss, with signed ESG telemetry co-produced at inference time. Cross-operator learning happens on aggregated gradients, never raw reservoir data, so competitive sensitivity is preserved. North-star: unplanned downtime avoided per well per year. Counter-metrics: reservoir engineer approval rate on recommended workovers, regulator audit-pass rate on methane filings.

Project 14 · Energy Transition · Utilities / ESG

Tideway: energy-transition capital allocation for an integrated utility.

A six-month, 10-year capital-allocation plan for a Canadian integrated utility, re-optimizing roughly $40B of capex against a three-objective frontier of carbon avoided, rate impact, and regulator sign-off.

Premise: a Canadian integrated utility (about $9B revenue, about 60 TWh/yr generation, about 3M customer accounts) is facing the energy-transition trilemma: a federal net-zero-electricity pathway (the finalized Clean Electricity Regulations, SOR/2024-263, set a 65 t CO₂/GWh performance standard for 2035 to 2049 and 0 t/GWh from 2050, and dropped the draft's low-run carve-out for peakers), rate-payer affordability constraints (politically and contractually capped), and an aging gas-peaker fleet that's both essential for reliability and a regulatory liability. Their current 10-year plan over-indexes on utility-scale solar because that's the cheapest unit cost in isolation; what it misses is the system cost of integration, which above a 20% variable-renewable share runs the effective system cost per MWh to about 2× the headline LCOE.

Tideway is a six-month strategic engagement: a 10-year capital allocation plan that re-optimizes against a three-objective frontier (MtCO₂ avoided, $/customer rate impact, regulator-signoff probability). Workstreams: scenario modeling across nine generation-mix futures, technology-bet sequencing (storage vs. small modular reactors (SMR) vs. demand-response vs. transmission), regulator-engagement strategy with the provincial energy board, and a stakeholder-impact sequencing plan that orders changes to land politically. North-star: cumulative MtCO₂ avoided per $B capex, optimized along the affordability constraint. Counter-metrics: grid-reliability index (cannot deteriorate vs. baseline), median rate-payer bill impact %.

Project 15 · Public Sector · Healthcare Operations

Helm: provincial wait-times reduction across a six-hospital cohort.

A nine-month operations engagement across six Ontario hospitals to bring admitted-patient ED length-of-stay toward the 8-hour provincial target by fixing downstream bed-flow, discharge governance, and triage.

Premise: Ontario's 90th-percentile wait for an inpatient bed from the ED reached 44 hours in 2024/25 against an 8-hour provincial target, with the average admitted patient spending 20.3 hours in the ED (Ontario Health / CIHI); worst-quartile sites materially worse; admitted-patient hallway-wait incidents have grown materially over three years. The Ministry of Health has set a tightened total-time-in-ED target as a measurable component of the next provincial budget cycle. The wait-time problem sits downstream in bed-flow, EMR-driven discharge friction, and a triage protocol that hasn't been re-tuned to current case-mix.

Helm is a nine-month operations engagement across six hospitals in two Ontario Health Teams (OHTs): patient-flow modeling at the asset level (ED → admission → discharge → bed-turn), discharge-readiness governance (medical reconciliation + transport + community-care intake), and a re-tuned triage protocol calibrated to 2025 case-mix. The work product is operational: visual-management boards in each ED, a daily flow huddle with clinical and ops leadership, and a discharge-readiness lookahead that moves bed turns earlier in the day. North-star: 90th-percentile total time-in-ED for admitted patients across the cohort, audited against the prior 90-day baseline. Counter-metric: 30-day readmission rate and a Maslach Burnout Inventory (MBI) burnout score, both held flat as guardrails.

Project 16 · Healthcare · Patient Experience

Patient Experience: what moves "would you recommend".

A key-driver analysis on 2.3M patient-experience survey responses, built from raw public data in SQL, finding which drivers move the recommendation rate and where to act first.

Premise: for a high-touch service, the useful question is which experience drivers actually move the recommendation rate, and where a limited budget should go first. The lowest-scoring dimension is often a weaker lever than it looks.

Built on the public CMS HCAHPS release, filtered to a 3,956-hospital analytic sample and 2.3M completed responses entirely in DuckDB and SQL, with no BI tool. A standardized key-driver regression ranks discharge information as the strongest driver of recommendation (23.4% of explained variance), shows quietness as both the lowest-scoring dimension and the weakest lever, flags early-warning states, and ties each recommendation to cited research (AHRQ, Doyle et al. 2013, Reichheld 2003). North-star: the share of patients who would definitely recommend.

Project 17 · Enterprise systems · CRM data quality

Concord: gating a CRM release on data quality.

A CRM / Opportunity-to-Engagement data-quality and release-testing console, built end to end in C# and SQL, that validates records against governed reference data, isolates what a release broke, triages the defects, and gates the release.

Premise: before a monthly CRM release goes live, someone has to answer whether it broke any records, which ones and why, and whether the open-defect backlog is healthy enough to promote. The useful artifact is a single call: ship it, watch it, or hold it.

Concord models a CRM pipeline in SQL (accounts, opportunities through the O2E stages, governed stage and currency reference data, and a defect register with a full OPEN to VALIDATED lifecycle). It holds two snapshots, a known-good baseline and a release candidate, runs seven data-quality rules against each, and treats any failure present in the candidate but not the baseline as a regression the release introduced. Each regression is triaged into the defect register, the release is gated OK / WATCH / HOLD by worst severity, and release-health KPIs (backlog, cycle time, validation rate) are reported. North-star: no high-severity regression reaches production. Rules map to the DAMA data-quality dimensions.

Project 18 · Manufacturing · MES & support

Takt: monitoring and supporting a shop floor.

A manufacturing-execution (MES) monitor built end to end in C# and SQL that tracks production KPIs and defect genealogy, then turns them into a per-line support verdict an engineer can act on.

Premise: at the end of a shift an MES has to answer which machine lost the most time and why, whether quality held, whether a work order has quietly stalled, and if a defect surfaces later, which units and which work order it traces back to.

Takt models the shop floor in SQL (work centers, work orders, a production-event log, and a defect table for genealogy), computes the core MES metrics in SQL (OEE decomposed into availability, performance and quality per Nakajima, a downtime Pareto, first-pass yield), then runs a support pass that turns the metrics into a per-work-center verdict of OK, WATCH or HOLD with the reason attached. Concepts map to SEMI E10/E79 equipment states and the ISA-95 MES reference model. North-star: no quality excursion or stalled work order goes unflagged.

Project 19 · Energy transition · Investment underwriting

Solstice: underwriting a solar-plus-storage asset.

A project-finance investment model for a 200 MW solar-plus-storage asset, built in Excel and Python: it constructs the 30-year cash flow, sculpts the debt, and returns the unlevered and levered numbers an investment committee needs to commit equity.

Premise: before an investment committee commits equity to a renewable asset, someone has to answer the returns (unlevered and levered), whether the debt service holds, what moves the equity IRR, and how much carbon the asset abates. The useful artifact is an auditable model plus a one-page recommendation.

Solstice builds a 30-year cash flow from generation, a 20-year contracted PPA with a merchant tail, a storage capacity payment, opex and capex; sizes and amortizes the debt with an annual DSCR; and computes the unlevered IRR and NPV, the levered equity IRR and MOIC, and the minimum DSCR. A sensitivity isolates the two swing factors, PPA price and capex, and a decarbonization read ties return to tonnes abated. On the base case: 9.5% unlevered IRR, 14.9% levered equity IRR, 1.9x minimum DSCR, and about 4.9 Mt CO2 abated over life. North-star: a disciplined, auditable underwrite an IC can act on.

Project 20 · Real estate · Acquisition underwriting

Cornerstone: underwriting a real-estate acquisition.

A real-estate acquisition model in Excel and Python: a 200-unit multifamily pro forma from rents through to NOI, a cap-rate valuation, amortizing debt with DSCR, and a hold-period exit, returning the levered and unlevered numbers an investment committee needs.

Premise: to underwrite a property acquisition, an analyst builds the pro forma, values it on a cap rate, sizes the debt, runs the hold and the exit, and answers the only question that matters: do the returns clear the bar, and what breaks them. The useful artifact is an auditable model plus a one-page recommendation.

Cornerstone builds NOI from gross rent through vacancy, other income, and opex; values entry and exit on cap rates; sizes and amortizes the loan with an annual DSCR and debt yield; and computes unlevered IRR, levered IRR, equity multiple, and cash-on-cash, with a sensitivity isolating exit cap and rent growth. On the base case, a $47M buy at a 6.0% going-in cap returns a 12.6% levered IRR, an 8.6% unlevered IRR, a 1.74x equity multiple, and a 1.50x minimum DSCR, with the exit held flat to entry so the return comes from NOI growth, not cap compression. North-star: a disciplined underwrite an IC can act on.

Project 21 · Energy · Solar performance engineering

Helios: operating a solar plant against its contract.

A utility-scale PV performance engine in Python: an expected-energy model for a 100 MW single-axis-tracker plant, the Performance Ratio and yield metrics owners and lenders track, an energy-loss waterfall, a capacity test that isolates excusable losses, and a fleet view that flags the plants drifting below guarantee.

Premise: a solar performance engineer is paid to hold a plant to the energy model it was financed on. The work is measuring delivered energy against expected, attributing every lost megawatt-hour to a cause, separating operator-controllable losses from excusable ones, and flagging the plants drifting below guarantee. The useful artifact is an auditable model plus a performance report.

Project 22 · Customer experience · Key-driver analytics

Lodestar: where an experience budget should actually go.

A key-driver model across a 3,956-hospital analytic sample and 2.3M completed CMS surveys. The dimension that attracts the budget turns out to be the weakest driver, and the strongest driver already scores highest.

Premise: hospitals read the patient-experience scorecard as a to-do list and start at the bottom. Across a 3,956-hospital analytic sample drawn from 2.3M completed CMS surveys, that instinct is expensive. Quietness at night scores lowest of the six experience dimensions at 55.8% top-box, which is why it attracts the budget, and it is tied for the weakest driver of whether a patient would recommend the hospital, at 12.1% of explained variance. The strongest driver is discharge information at 23.4%, and it already scores highest at 85.9%. HCAHPS feeds CMS Hospital Value-Based Purchasing, so the misallocation lands on a payment line rather than a reputation.

Lodestar is an eight-week diagnostic, run before any improvement budget is committed. The six dimensions are strongly correlated, so raw regression coefficients are unstable and can flip sign; the model uses Johnson relative weights, which split the R-squared (0.57) into non-negative shares per driver. The engagement builds the driver model on the client's own survey file, maps importance against current score, separates a local operations problem from a national pattern, then costs the discharge-information intervention and the measurement behind it. North-star deliverable: a priority list the board can fund, and a formal decision to stop the programme it would otherwise have started.

Project 23 · Capital Markets · Applied AI

Corroborate: an event record you can check rather than trust.

A capital-markets event timeline holding 63,505 macro, markets and micro events, with grounded search and per-company recommendations that trace back to the filing or release they came from.

Premise: research tools answer confidently and leave you unable to check the answer. Corroborate stores the primary record first, then answers only from it. Every event carries its source, its timestamp and its filing type, so a claim can be followed back to a SEC filing, a Federal Reserve release or a FRED series rather than accepted.

Built on Cloudflare Workers with D1, Vectorize and Workers AI, ingesting on a 15-minute cron across SEC EDGAR, FRED, Federal Reserve releases and market news. Filings and macro history reach five years; news reaches March 2022, where the only free archive begins. Ratings are derived from a published scoring rubric rather than asserted, and the last rating is saved with its timestamp so the tool still answers when the daily model allowance is spent.