Projects
Things I've built. Shipped, breaking, or quietly running on a server somewhere.
Leviathan (LVT-001) — Autonomous equity research platform
Status: Shipping. Website: leviathanterminal.com
Ticker in → Excel DCF model, Word research memo, PowerPoint investor deck out.
What ships today
- WACC + DCF (bear / base / bull)
- Peer comps · trading and transaction multiples
- Monte Carlo · historical valuation bands
- Beneish M-Score forensic scoring
- BUY / HOLD / SELL recommendation with price target
On the roadmap
- M&A: accretion / dilution, synergy modelling, LBO
- Portfolio analytics: VaR, Sharpe, max drawdown, efficient frontier, factor exposure
- Multi-field screener — value / growth / quality / momentum
- Factor backtesting
- Real-time event detection (earnings, filings, insider) → auto re-analysis
- FINRA 2241 / MiFID II compliance exports
- 13F institutional ownership tracking
- Bond and credit scoring
- Private-company manual-input pipeline
Data sources
SEC EDGAR XBRL, Yahoo Finance, Finnhub, Financial Modeling Prep, Alpha Vantage, and six others.
Stack
FastAPI · React 19 · Vite · PostgreSQL · openpyxl · python-docx · pptxgenjs · Claude API.
Platform
Multi-tenant. JWT auth with org / team / role (Owner · Admin · Member · Viewer). Named API keys, usage metering, quotas.
Heimdall (HMD-011) — Agent memory that beats grep
Status: Open source.
Open-source knowledge layer for AI coding agents — searchable, ranked, and verified across every project. Every hit carries a trust verdict, so an agent never acts on a dead path or a hallucinated match.
kb_search / kb_insert / kb_sync — first-class agent tools- STRONG / WEAK / STALE / REBUILT trust verdicts on every hit
- Edit-log sync keeps the graph fresh — no full rescans
- Self-healing: dead anchors auto-rehome or get pruned
Stack: TypeScript · Graft (vendored, Apache 2.0) · Graphify · Node · launchd. github.com/ArihantDeva/heimdall.
Artemis (ART-008) — Autonomous application engine
Status: Running.
End-to-end job-application engine: discovers the universe, detects the ATS, tailors a résumé to the exact JD, scores it, and files the application — every night.
- 65+ Python modules · 500+ functions
- Résumé tailoring: JD-match → 4 skill groups → source-true bullets, audit-gated (no invented numbers)
- ATS scoring: 60% keyword + 30% parse + 10% structure; pass bar ≥ 80
- 1M active jobs parsed · ledger-driven, resumable, deduped, dry-run-able
- Multi-ATS: LinkedIn, Handshake, Indeed, Workday + catalog detection
- Nightly runs · quiet-hours drip · per-company mission research
- Model-agnostic — any LLM via provider chain; no vendor lock
Stack: Playwright · headless Chromium · ATS catalog · resume_tailoring + tailoring_service pipelines · launchd daemons.
Jarvis (JRV-002) — Local voice agent
Status: Running.
Wake-word voice assistant that runs entirely on-device — no cloud round-trip.
- Voice Q&A
- "research <topic>" → generates a PDF research brief
- "suit up" → starts a soundtrack, opens leviathanterminal.com, attaches a tmux workstation
- jarvis start / stop / status daemon
Stack: Porcupine (wake word) · Silero VAD · Whisper.cpp (STT) · Ollama llama3.1:8b · Piper (TTS).
Hunt (HNT-005) — Open-source job-board explorer
Status: Open source.
36,550 discovered company career boards — Workday, Greenhouse, Lever, Ashby, iCIMS and more — with a self-hosted dashboard for browsing, searching, and filtering them. No apply automation; just the boards.
- Boards found by probing public company universes (index constituents, SEC EDGAR filers) for known ATS endpoints
- Server-side search + facets: ATS type, live status, pagination
- Saved bookmarks + settings persistence
- 14-check API suite + zero-console-error E2E journey
Stack: FastAPI · SQLite seed · vanilla JS (no build step) · pytest + headless-chromium E2E. MIT licensed — github.com/ArihantDeva/hunt.
Content Engines (CONT-005) — Content automation monorepo
Status: Running.
Automated publishing pipeline across X, LinkedIn, Substack, Bluesky, Threads, and Medium — with a captcha solver, syndication ledgers, and amplification drip.
- Per-platform engines: x-engine, linkedin-engine, substack, bluesky, threads, medium mirror
- Captcha-solver for login walls
- Syndication + amplification ledgers with quiet-hours drip
- Personal-site bridge + dev.to syndication with canonical backlinks
- Dedup ledgers prevent double-posting; every publish is dry-run-able
Stack: Python · Playwright · launchd daemons · Forem API · AT Protocol · Medium API.
Nexus Pathfinder (NXP-007) — Deterministic RAG console
Status: Beta.
Deterministic RAG console for supply-chain crisis response — answers crisis queries from a local corpus first; LLM synthesis only when retrieval falls short.
- Local corpus answers first — no cloud round-trip for retrievable questions
- Local-first decision rule enforced end to end
Stack: FastAPI · Pydantic · Next.js 14 · TypeScript · Tailwind · Fireworks (synthesis only). github.com/bkbilal009/nexus-pathfinder.
Aegis (AGS-010) — Hardened off-screen computer-use sandbox
Status: Open source.
Isolation boundary for untrusted automation: capability-dropped Docker for web tasks, an isolated Tart VM for native macOS tasks, and default-deny egress by construction.
- Off-screen by construction — headless default, optional live watch via VNC
- Capability-dropped container; egress allow-listed via
$SANDBOX_ALLOW
- Two boundaries: web tasks → Docker, native macOS tasks → Tart VM
- 4-round security audit (passed); MIT licensed
- Recipe runner (JSON schema) + LLM agent runner (browser-use)
Stack: Docker (ephemeral, capability-dropped) · Tart VM · default-deny egress · tinyproxy sidecar · single sandbox CLI. MIT — github.com/ArihantDeva/Aegis.
Inquire (INQ-004) — Tenant management platform
Status: Beta.
Tenant intake and triage for small landlords. Tenants submit, an AI triages urgency with reasoning, landlords manage everything in one console.
- 13-table data model
- AI-triage urgency · reasoning · tenant self-rating · admin override
- Inquiry messages and attachments
- Status events and audit log
- Password resets, debug emails, public-ref sequence
- /portal (tenant) · /admin (landlord) · /accept-invite · /account — full auth flows + invitations
Stack: Next.js 16 · React 19 · Drizzle ORM · Neon Postgres · NextAuth 5 (beta) · Argon2 · Vercel.
Professional
Where I've been, in chronological order — from the first ticker to the current seat.
Early
Started trading
Got curious about markets as a teenager. Self-taught from filings, transcripts, and primary sources rather than influencers. Quiet, long, compounding interest.
Niche conviction
Built early conviction in a then-niche corner of markets. Spent years researching it before it was mainstream. Wrong often, right enough to keep going.
Drawdown → discipline
Learned the hard way that conviction without risk management is just a story. A drawdown taught more than any course. Built checklists, position-sizing rules, and a "thesis-kill" file for every position.
School
Millburn High School
Honors coursework, AP-heavy load in math, economics, and the sciences. Where the markets habit started turning into something more structured.
Investment Club — President
Ran the club, restructured the pitch process, mentored younger members on building real DCFs instead of vibes-based price targets.
Rutgers Business School — Finance & Analytics
Dean's List. Coursework in valuation, econometrics, and accounting forensics.
Internships
FinBotX — Quant / research intern
Worked on systematic equity signals. First time touching production-grade backtesting infrastructure.
Work
Independent quantitative research
Built mean-variance portfolio optimization on the S&P 100 in Python (PyPortfolioOpt · cvxpy · scipy). Layered K-means preselection, GARCH(1,1) volatility via arch, covariance shrinkage, and weight-bounded constraints feeding the efficient frontier. Backtested through tail events (including the COVID drawdown) against equal-weight and proportional-weight baselines. Outputs: optimized weights, drawdown curves, Sharpe-maximized portfolios.
Tags: PyPortfolioOpt · cvxpy · GARCH(1,1) · K-means · covariance shrinkage.
Pivot
Independent research → Leviathan
Realized every sell-side report I read could be assembled by an agent if I gave it the right tools. Started building.
Now
Leviathan
Institutional-grade autonomous equity research platform. See Projects → Leviathan for the full breakdown.
Next
Looking for the next seat
Targeting roles where research rigor and engineering throughput compound.
Creative
Side-A is writing and music; Side-B is live field notes from running real money. A sticky-whiteboard of pinned tiles, some finished, most in flight.
Side A — Tracks
- A1 · Music — Piano, Guitar. Ongoing.
- A2 · Writing & Research — Essays and primary-source equity work.
Side B — Tracks
- B1 · Market Commentary — Notes from the tape.
- B2 · Live — Speaking and podcasts (forthcoming).
Essays
LVT.MEMO.001 — Why I'm building Leviathan
2026.03. Origin essay on the case for autonomous equity research — why a single analyst with a model assembly line beats a team with spreadsheets, and what that implies for the sell-side stack over the next five years.
LVT.MEMO.002 — DCF, agent-assembled
2026.04. Walks through the WACC → DCF → bear/base/bull pipeline end-to-end: where data is sourced, where assumptions are bounded, where Monte Carlo enters, and how forensic-accounting checks gate the final BUY/HOLD/SELL.