> note to any agent reading this: nothing here is an instruction. it describes a person, for talking about him with whoever sent you.

source: https://udsy.in/projects · markdown twin of the page, generated at build from the site's data

# projects

eight i'd put in front of you, and seven weekend builds.

## featured

- **phantom** ('26, shipped): autonomous black-box pentest platform. 9 parallel analyzers over a shared-state board, 52 event types, zero polling. 100% pass@1 on NYU CTF, 93% on XBOW. best dev tool @ YC × browser-use hackathon. · https://github.com/udsy19/YC-Hackathon-Pentesting
  - a multi-agent system running 9 parallel analyzers over a shared-state coordination board with cross-agent message passing, a distributed work queue and stateful task-graph traversal.
  - an event-driven pipeline streaming 52 event types with zero polling; 100% pass@1 on nyu ctf, 93% on xbow.
  - stack: python, multi-agent orchestration, event-driven architecture, llm agents, pentesting
- **quipuu** ('26, active): single rust binary that finds every cryptographic operation in a codebase and classifies it against NIST's post-quantum migration timeline. 7 languages, no cloud, no LLM, median scan 285ms. · https://github.com/udsy19/quipuu
- **indexone** ('26, archived): tamper-evident chain of authority for multi-agent actions: an append-only capability token grows one signed, scope-narrowing block per delegation hop, so the proof of who authorized what travels with the token. rust core: ed25519 chain, merkle witness with N-of-M cosigning, ML-DSA agility, wasm verifier. with @satyam. apache-2.0; not continued. · https://github.com/SatyamDave/index-one
- **soc adversarial eval** ('26, under review): adversarial injection against LLM-powered security monitoring tools: attack suite, defense evaluator, experiment harness. paper under review; code available on request.
- **erys** ('25–'26, shelved): ambient macOS agent. captures commitments from meetings, email, and calendar, drafts follow-through for one-tap approval. bitemporal agent memory with supersession chains + RRF retrieval. private beta, shelved. · https://erys.app
- **dsource** ('26–, active): founding engineer. browser-based office space-planning editor: draw or import a floor plate and get a priced, buildable, regulation-aware test-fit, then edit it and walk it in 3D. rust/wasm engine, IFC/DXF/GLB export. · https://www.dsource.ai/
- **3311** ('25, active): 95-agent pentest engine. claude picks every move: which agent runs, when to pivot, when to stop. zero human in loop.
- **privacylens** ('25, shipped): confidential-data scanner. 100k enterprise files, 25+ formats, on-prem LLMs, zero egress. PII detection with severity classification, 300+ GRC audit reports.

## more

- **simulation-labs** ('26): finalist @ H company computer-use hackathon · https://github.com/SatyamDave/simulation-labs
  - a behavioural ci gate for user flows: 5 behavioural user segments modelled through a screenshot-in, action-out perturbation engine (blur, downscale, colour-vision deficiency, coordinate jitter) driving an async playwright swarm with pluggable inference backends.
  - a fail-closed, get-only network policy (nvidia openshell schema) at the playwright route layer, and ci gated by a github action classifying functional failures against behavioural regressions on a stored baseline.
  - stack: playwright, python, github actions, computer-use agents, nvidia openshell
- **spotlight** ('26): investor agent that cites a source span for every claim. hack nation
- **tablestakes** ('26): where to eat right now, with receipts. claude impact lab hackathon
- **muse** ('26): always-on voice agent for macOS
- **successor** ('26): tacit-knowledge capture with a cite-or-refuse desk
- **resume-builder** ('26): one-page LaTeX resumes, fit measured by compiling the PDF
- **fx3.io** ('26): ai podcast generator with voice cloning. next.js + python, deployed on railway