Full-stack and AI/ML engineer. Most of what I build sits under agents rather than inside them: settlement rails, permission layers, adjudication, the machinery that has to hold when nobody is watching.
I work upstream too: 374 pull requests and 58 commits to Nous Research's Hermes Agent, plus its plugin repos. TUI, gateway, tool layer, desktop runtimes, state stores, adapters.
| Working on | Role | State |
|---|---|---|
| PayMate — non-custodial USDC billing rail for agents | author | live on GOAT mainnet |
| MediVault — private AI health vault on 0G | author | live, ERC-7857 contract on mainnet |
| Hermes Agent — agent runtime and plugin catalog | upstream contributor | 374 PRs · 58 commits |
| GenLayer adjudication — evidence-first verdict contracts | author | 4 protocols shipped |
Hermes is the agent that grows with you — a real agent runtime with desktop apps, gateways, channels, sandboxes, memory and a plugin catalog. I spend my time in the failure modes: a gateway that won't release a stale PID, a Windows profile path with non-ASCII characters, a WAL generation unlinked by a short-lived reader, an OAuth login that a single refresh can revive.
Shipped upstream:
| PR | What it fixes |
|---|---|
hermes-agent#140922 |
gateway: stamps an unproven --replace refusal as startup_failed so the next start wins |
hermes-official-plugins#61 |
directsdk: carries the login identity into the child env, captures native stderr |
hermes-labs-plugins#3 |
record: an autocomplete token inside a list is still a secret field |
The patches themselves — all 58, every one in hermes-agent, spread over 20+ topic branches. 36 are already on main; the rest are in open PRs or awaiting rebase:
| Date | Commit |
|---|---|
| 2026-10-21 | fix(gateway): log every missed liveness probe and dump the loop's task state |
| 2026-10-21 | fix(gateway): a retiring workspace takes no new work (#134889 TUI-27) |
| 2026-10-21 | fix(browser): an unbound thread resolves no configured CDP endpoint (#134889 GW-26) |
| 2026-10-21 | fix(checkpoints): pin literal pathspec mode for plan-owned git selections |
| 2026-10-21 | fix(checkpoints): drop oversize files from the index by literal pathspec |
| 2026-10-21 | fix(agent): serialize stdio installer transactions on the shared install lock |
| 2026-10-20 | fix(tools): do not drop a tool whose probe answered at the deadline |
| 2026-10-20 | fix(tools): bound a check_fn probe so a wedged availability check cannot hang the turn |
| 2026-10-20 | fix(approval): scope the git force-push patterns to the push's own command segment |
| 2026-10-19 | fix(stdio): never stack Hermes stdio proxies across the agent-build cycle (#139687) |
| 2026-10-19 | fix(prompt): render toolset-gated guidance in the volatile tier (#137307) |
| 2026-10-19 | test: tolerate bases without the redacting stdio wrapper in the proxy-depth helper |
| 2026-10-18 | fix(gateway): park a unit on unprovable --replace ownership instead of restart-looping |
| 2026-10-18 | fix(gateway): let the startup_failed stamp clear the way instead of parking the unit |
| 2026-10-18 | fix(gateway): make the ownership-refusal stamp durable before the hard exit |
| 2026-10-07 | fix(agent): gate the 20-image eviction trigger on the request's dimensions |
| 2026-10-07 | fix(anthropic): give the wire image-eviction pass the 24 MB request budget |
| 2026-10-02 | fix(state): key display identity of tool rows on stable tool identity, not payload bytes (#117750) |
| 2026-09-29 | fix(cron): keep a sticky last_failure stamp so a healed run stays visible |
| 2026-09-27 | fix(auth): treat nous as a single-use refresh pool provider |
| 2026-09-27 | fix(memory): migration note must not imply config.yaml provider keys still apply |
| 2026-09-26 | fix(backup): percent-encode path in _query_ro_sqlite URI |
| 2026-09-14 | fix(whatsapp): adapter reads WHATSAPP_GROUP_ALLOWED_USERS for the group allowlist (#72529) |
| 2026-09-14 | fix(discord): dispatch-side liveness dimension detects an ACKing-but-deaf gateway socket (#109521) |
| 2026-09-13 | fix(agent): same-model review fork inherits the parent's resolved cache scope (#109964) |
| 2026-09-13 | fix(mcp): skip dynamic tool refresh when the server session is already gone (#109824) |
| 2026-09-13 | fix(agent): mirror Claude Code OAuth refresh into the macOS Keychain (#98334) |
| 2026-09-13 | fix(discord): liveness probe gains a dispatch-side dimension (#109521) |
| 2026-09-12 | fix(desktop): do not fail build if get-windows is missing on any platform |
| 2026-09-11 | fix(kanban): preserve sticky block on tasks created with initial_status=blocked (#107398) |
| 2026-09-11 | fix(tui_gateway): handle deleted profiles gracefully in _response_profile_name (#107829) |
| 2026-09-10 | run: quiesce cron/deferred workers before DB close |
| 2026-09-09 | fix(serve): reserve ProcessLookupError for known missing process signals |
| 2026-09-09 | fix(serve): ensure parent-death watchdog detects exited Desktop parent (#80204) |
| 2026-09-03 | style: separate vitest import from local imports (perfectionist/sort-imports) |
| 2026-09-03 | fix(desktop): assistant rows whose text persisted to sidecars no longer vanish on rehydrate (#68321) |
| 2026-09-03 | lint: fix import ordering in chat-messages.reasoning-survival.test.ts |
| 2026-09-02 | fix(recovery): make the printed salvage command satisfy the real CLI contract |
| 2026-09-02 | fix(recovery): stop pointing sqlite3 .recover guidance at the live state.db |
| 2026-09-01 | refactor(gateway): drop constant conditional in rescue helper |
| 2026-09-01 | fix(update): break the cron-update three-way restart deadlock (#100179) |
| 2026-09-01 | fix(gateway): rescue orphaned FIFO overflow when session goes idle (#99882) |
| 2026-08-31 | fix(preview): use proper input events for React/contenteditable/sumbit in drive_preview type |
| 2026-08-30 | fix(agent): exempt :cloud GLM models from stop->length truncation rewrite |
| 2026-08-30 | fix(desktop): honour arrowed slash-picker highlight on bare '/' query |
| 2026-08-30 | test(agent): split GLM cloud/local stop-continuation cases |
| 2026-08-29 | fix(cron): accept bare duration units like 'hour' in schedule parsing |
| 2026-08-29 | fix(cron): mention bare units in the duration parse error |
Open and in review: gateway restart paths · stale heartbeat requeues · execute_code kernel memory caps · wedged check_fn probes · bounded inbound media download · local-time timestamps in Home Assistant.
Surfaces touched: TUI · gateway lifecycle · tool execution and output spilling · cron and heartbeats · desktop (Windows/macOS) · MCP · ACP · SQLite state, WAL, backups · compression and checkpointing · installers · Discord / WhatsApp / xchat adapters · DirectSDK · Spotify · Home Assistant.
Also authored 10 Hermes skills published as standalone repos — daily briefing, release radar, disk guard, IP watchdog, uptime digest, cert expiry, price digest, battery, dead link, folder size.
Also shipped: JanFix (civic reporting, zero login, ward routing) · Monaco CLOB dashboard (Sei order book) · SnakeCase (34-lesson Python curriculum) · ZeroSense (ZK robot proofs + Stellar micropayments) · Reso (ZK compliance engine) · Atropos fork (LLM RL environments).
Agents can chat, code and research. They cannot collect money. Invoices are PDFs, chasing is manual, and nothing exists for a per-call micropayment. Freelancers, meanwhile, have no portable proof that they finished anything.
PayMate closes both sides: a freelancer or an agent creates an invoice or a paywalled endpoint, the client settles in real USDC directly to the freelancer's wallet, and every settlement mints a portable on-chain reputation credential. No custody of principal, no API-key gatekeeping.
Live: paymateagent.xyz · Docs: paymateagent.xyz/docs · Developers: paymateagent.xyz/developers
| Piece | Mechanism |
|---|---|
| Settlement | USDC on GOAT mainnet (chain 2345), client → freelancer wallet, verified on-chain |
| Agent billing | POST /api/agent-billing mints per-call x402 challenges; replay-guarded unlock |
| Reputation | ERC-8004 soulbound credential per settlement — travels with the agent |
| Escrow | YieldEscrow locks funds; PR merge or AI verdict releases them |
| Disputes | Mistral arbitrator rules against the original scope: pay / refund / split |
| Cross-chain | client signs one tx on any of 39 networks; relayer converts and settles in USDC |
| Zero-custody rail | swap → approve → bridge, DOGEB lands straight in the freelancer's wallet |
| Self-refill | BSC deposits pivot, bridge and DEX-swap into the payout pool automatically |
| Operator agent | audits the ledger, checks liquidity runway, retries failed swaps under hard USD caps |
Deployed and verifiable — YieldEscrow 0xEbe3BE16d7fd69268BcADdB6DB25C60dFff302e0, PayMateReputation 0x48B9B6BB1C4Eb4b6911D05dfd3E87F8be8a67603, PayMateTreasury 0x59C4c64DF838f23b8dA77A95c18BC3520D71e25B on GOAT mainnet. Settlement counts, fee take and every hop are on the public growth and treasury pages — read from the API, not asserted in a deck.
The part I actually care about: nothing moves on trust. Bridge and DEX paths ship fail-closed behind verification flags until a real-money probe passes, the relayer's ledger is idempotent, and the AI operator can prioritise but never exceed a code-enforced cap.
Smaller, stranger builds where the constraint is the point — every one of them a working repo, not a notebook:
quantum-entangled-transformer (attention heads as entangled states) · cellular-automata-transformer · superconducting-super-computation (Josephson-junction SNN simulator) · seismic-anomaly-detection-snn · spiking-weather-net · topological-data-homology-nets · thermodynamic-neural-annealing · viscous-fluid-neural-flow · swarm-drug-molecule-designer · symbiotic-multi-agent-rlaif.
Everything I actually work in, no padding.
Open to work on agent infrastructure, settlement and adjudication, and hard reliability problems in long-running systems.
