Skip to content
View pham-canh's full-sized avatar

Block or report pham-canh

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
pham-canh/README.md

Phạm Cảnh

Software Engineer | AI Agents & Automation

I work on AI agents, automation, and backend systems.

I'm interested in building AI systems that can handle tasks, work together, and check their own results.

Recently, I've been working on multi-agent workflows with a focus on:

  • Building agents with separate roles and responsibilities.
  • Automating tasks from start to finish.
  • Checking results against actual files and code.
  • Handling errors, retries, and task recovery.

I'm always learning and building new things.

Current build: Trip Relay, two travel agentic websites with a per-origin consent gate, for the Open Agentic Web Challenge.

Based in Ho Chi Minh City, Vietnam. Open to remote roles.

Selected work

opencode-harness-kit — model-agnostic harness for OpenCode. Capability profiles, context pruning, model failover, and a verify command that checks the installed config against the profile. MIT, CI on every push, published on npm as opencode-harness-kit.

pocketful-factory — three coding-agent seats build a wallet service from a written spec, with an independent evaluator that re-derives every figure and fails when a document disagrees. 4/4 stages, 36 claims checked by verify_claims.py, all 121 room messages written by agent seats.

voicecare-scheduler — voice booking for elderly clinic patients: speak, hear the appointment read back, and nothing is booked without a spoken yes. 28 tests, per-stage latency, and the UI names the provider that actually answered. ▶ 42s capture

eurorwa — research project on independently verifiable snapshots for tokenized money-market funds: hash the snapshot, sign it, publish the attestation on-chain, and let anyone re-hash and check the signer with no credentials.

Competition entries

Event Project What it shows
BAND Dark Factory — WeAreDevelopers, lablab.ai (2026) pocketful-factory multi-agent factory, independent evaluation, 4/4 stages, offline-built service
Nebius × NVIDIA Global AI Hackathon — Devpost (Oct 2026) voicecare-scheduler voice agent with a hard consent gate; Nemotron NLU with labelled fallbacks
Open Agentic Web Challenge — Cognizant × Mozilla, Devpost (2026-27) trip-relay airline and hotel browser agents, per-origin consent before contact, closed peer policy
AssemblyAI Voice Agent Hackathon — lablab.ai (2026) assemblyai-voice-agent (archived) realtime STT → NLU booking loop; no booking without a spoken YES
Vultr Agent Rush (2026) consult-homeopathic (archived) patient intake with a human doctor in the loop

Archived, not pinned: all other hackathon builds are read-only — dark-factory (superseded by pocketful-factory), blast-radius-zero, amd-web-retrieval-agent. Archived entries stay public so their links keep working; the four pinned repos are the current portfolio.

Research

Separate from the work above, eurorwa is a research project, not a product. It studies how public data for tokenized money-market funds can be collected and then independently verified — a daily snapshot is hashed (keccak-256), signed with a dedicated key, and the signature published on-chain, so anyone can re-hash the data and check it against the attestation. MIT licensed, no paywall, no API key, no hosted service.

failed-strategies keeps backtested strategies that lost money, with post-mortems. Published on purpose: knowing why an approach failed is more useful than hiding it.

Listing the research apart is the point: the agent work above is what I built to ship; these are questions I am still working on.

Tech stack

Python  ·  TypeScript / JavaScript  ·  Bash
Node  ·  Docker  ·  GitHub Actions
PostgreSQL  ·  Redis  ·  AWS  ·  OpenTelemetry
LLM providers: NVIDIA Nemotron (Nebius) · Gemini · OpenAI
LLM routers: Mozilla Otari
STT/TTS: AssemblyAI · Edge TTS

What I care about

Claim it, then make it checkable. A fallback that silently downgrades the system is worse than a loud failure, and a README number that the app contradicts is a bug I want to hear about.


Open to contract work on agent systems, LLM integration, and automation.

Or open an issue on any repo above.

Pinned Loading

  1. eurorwa eurorwa Public

    Research: independently verifiable snapshots for tokenized money-market funds — hash, sign, on-chain attestation, re-verify with no credentials.

    TypeScript

  2. opencode-harness-kit opencode-harness-kit Public

    A model-aware harness for OpenCode: bring your own model, compile it into job lanes with measured baselines, privacy-aware routing and a verifiable install.

    JavaScript

  3. pocketful-factory pocketful-factory Public

    Dark Factory (BAND x WeAreDevelopers): three coding-agent seats in one room build a wallet service to spec, one reviews its own work and can veto it, and every figure is checked against the evidence

    Python

  4. voicecare-scheduler voicecare-scheduler Public

    Voice booking for elderly clinic patients — speak to book, nothing books without a spoken yes. NLU on NVIDIA Nemotron via Nebius Token Factory.

    Python