ankit.kapoor /

Engineering & Specialized Systems

More Projects

Working systems outside the core strategic cases.

Alongside the five decision systems featured on the portfolio, I build domain-specific engines, computational experiments, and valuation workbenches. Each project operates on a live deployed demo and open-source codebase, holding the same commitment to visible assumptions, reproducible pipelines, and mathematical rigor.

Healthcare AI · Clinical Decision SystemsLive Prototype · Open Source

OrthoCode AI

How do we make clinical AI verifiable by auditors and clinicians?

An evidence-first orthopedic medical-coding platform. Operates on a normalized, versioned codebook release and preserves page-level chart evidence for every proposed coding line.

Key Architecture & Mechanics

  • Normalized codebook releases: Ingestion and validation of 11,525+ licensed 2026 CPT codes and 4.49M directional NCCI PTP edits with cryptographic raw-file SHA-256 provenance.
  • Evidence-linked fact extraction: De-identified operative PDF notes parsed via OpenAI Structured Outputs (store=false) to produce bounded, page-anchored clinical fact candidates.
  • Calibrated decision graph: TypeSafe JEV decision engine with calibrated green/yellow/red confidence thresholds, failing closed to mandatory human review on uncertain cases.
  • Deterministic rule verification: Python validation for active-code service dates, NCCI PTP unbundling, MUE units, add-on dependencies, and HCPCS administered unit arithmetic.
Python 3.12FastAPINext.jsPostgreSQL 16RedisDockerTypeSafe JEVOpenAI Structured Outputs
Computational Neuroscience · Reservoir ComputingLive Prototype · Open Source

FlyCast

Can an evolved biological nervous system serve as a physical recurrent computer?

Turns the experimentally reconstructed MaleCNS v1.0 Drosophila melanogaster connectome into a fixed recurrent computational reservoir to test whether real biological wiring provides forecasting utility.

Key Architecture & Mechanics

  • Biological substrate: 166,483 identified neurons and 25.58M directed synaptic edges modeled as a high-performance Compressed Sparse Row (CSR) matrix with presynaptic-to-postsynaptic orientation.
  • Recurrent Echo State dynamics: Leaky rate dynamics running over frozen biological wiring with log-scaled synaptic contacts and incoming L1 normalization.
  • Virtual state probes: 4,096 virtual electrodes sample internal network state trajectories to train a lightweight multi-horizon linear Ridge regression readout.
  • Scientific honesty: Evaluated with zero data leakage against held-out test data and benchmarked against autoregressive baselines, persistence, and weight-shuffled controls.
PythonNumPySciPy (Sparse CSR)Scikit-LearnReactNext.jsMaleCNS v1.0 Connectome
Quantitative Finance · Asset ValuationLive Prototype · Open Source

Catalog Underwriter

Underwrite a music catalog using only what the public can see.

An interactive public-data workbench for reconstructing music catalogs and stress-testing valuations—separating observed consumption, modeled decay, financial assumptions, and discounted cash flow.

Key Architecture & Mechanics

  • Public evidence assembly: Live MusicBrainz artist identity resolution paired with optional Last.fm relative scrobble demand and YouTube view metrics.
  • Demonstration underwrite: Stored snapshot for Kendrick Lamar allowing full catalog valuation and scenario exploration without external API dependencies.
  • Interactive DCF modeling: User-editable rights shares, decay profiles, capital costs, bear/base/bull cases, and two-dimensional discount vs. growth sensitivity matrices.
  • Break the Deal diagnostics: Deterministic stress testing pinpointing the exact combinations of decay acceleration, concentration risk, and discount rates that impair investment returns.
React 19TypeScriptTailwind CSSRechartsMusicBrainz APILast.fm APIYouTube Data API
Equity Research · Company ValuationLive Prototype · Open Source

ValueLens

What is a stock price betting on?

A filing-based valuation workbench that rebuilds a company's cash flows from its SEC filings and reads today's share price three ways—the growth the price requires, what analysts expect, and what the fundamentals support.

Key Architecture & Mechanics

  • Sourced filing extraction: Revenue, operating income, R&D, capex, cash, debt, and shares pulled from 10-K and 10-Q filings with provenance on every figure; lines a company does not report are labeled rather than estimated.
  • Three readings of every price: Market-implied five-year revenue growth, an analyst-consensus case net of buybacks or dilution, and a fundamental FCFF DCF with R&D capitalized as investment.
  • Measured value ranges: 1,000 scenarios drawn from measured forecast errors produce an 80% value range, calibrated on held-out data at 80.8% for revenue and 78.7% for margin.
  • Backtested, not asserted: Default rules tested point-in-time on a 155-issuer SEC archive cut five-year revenue forecast error from 40% to 30%, while return backtests showing no demonstrated edge are reported openly.
JavaScriptCloudflare WorkersSEC EDGAR / XBRLFCFF DCFMonte Carlo SimulationOpenAI API
Healthcare Billing · Price TransparencyLive Prototype · Open Source

Bill Buster

Which hospital charges are worth questioning, and how do you ask?

Turns a coded outpatient facility bill into an evidence-backed inquiry letter and phone checklist. PDF reading, billing checks, and letter generation all run in the browser, and findings are framed as questions to ask rather than conclusions.

Key Architecture & Mechanics

  • Deterministic billing checks: Possible duplicates, CMS outpatient MUE unit limits, and Q3 2026 hospital NCCI procedure-to-procedure edits, applied with indicator, modifier, provider, and effective-date logic.
  • Hospital price context: Charges compared against Keck Hospital of USC's published price file, with a field-specific reason given whenever a price cannot be matched.
  • Cited inquiry packet: A first-person letter and phone checklist exported as PDF, with every citation linked to its source record and a conditional reduced-bill scenario labeled as not approved.
  • Private by design: Bills stay on the device; optional AI drafting via a Cloudflare Worker sends only confirmed findings, and figures and citations are always inserted by code.
JavaScriptPDF.js (unpdf)pdf-libCloudflare WorkersOpenAI Responses APICMS NCCI & MUE Data
Political Psychology · Survey MeasurementLive Prototype · Open Source

Ideology Inventory

Where do your political views actually sit?

A research-informed political-orientation inventory that places respondents on separate economic, social, and state-authority scales, adds short climate, trade, and populism scales, profiles six moral foundations, and checks how far to trust the answers.

Key Architecture & Mechanics

  • Multidimensional measurement: Three 12-item core scales grounded in the political-psychology literature, plus six MFQ-2 moral foundations scored relative to the respondent's own average.
  • Noise controls: Every ideology statement is matched with an opposite-keyed partner so acquiescence cancels out, and "no opinion" widens the uncertainty band instead of pulling scores to zero.
  • Response-quality grading: Seven checks—attention, implausible statement, contradictory pairs, agreement tendency, response variety, reading speed, and no-opinion count—combine into a High, Moderate, or Low grade.
  • Honest benchmarking: Only questions copied word for word from ANES 2024 and CES 2024 are compared with weighted U.S. distributions; consented anonymous responses feed a reliability, IRT, and test–retest pipeline in R.
TypeScriptViteCloudflare WorkersCloudflare D1TurnstileR (IRT / CFA)Python

Open Source & Code

Explore further work on GitHub

Beyond the five core decision cases and the systems above, my GitHub profile hosts ongoing code, open-source tools, and research experiments across machine learning, system design, and quantitative strategy.