At Mantiq I write Rust across an unusually wide range — firmware on the microcontrollers inside our drones and boats, the backend that talks to them, the frontend that drives the ERP. Same language, three very different worlds.
After hours, AI is the obsession. Specifically, the parts of agent systems everyone skips because the model gets all the attention — context engineering, model routing, the moment a human needs to say "send it." Epitech, promo 2026.
coda · python agent harness
A runtime with a first-class ctx namespace for context engineering, line-level observability, and a real human-in-the-loop UI. The agent reads your inbox, drafts a reply, and stops before sending. Every line of working Python is captured as a structured event — replayable when it does something dumb, savable as a skill when it does something useful.
Self-debugged its own failure mode once. Handed a fresh agent a past failed trace; it identified a bug I'd missed and proposed the right patch. The patch landed.
forge · tool-aware model router (rust)
Why use one model for an entire agent loop when every tool boundary is a natural swap point? Cheap models digest tool output. Smart models plan. Vision models only when an image was just produced. Local models for PII.
--after-tool web_search:claude-haiku-4-5
--after-tool execute_sql:openai/ft:gpt-4o-mini:org:sql-v3
--after-tool screenshot:gemini/gemini-2.5-flash
Plus a content-addressed run graph so you can audit which downgrades are actually safe against your real recorded traces.
agent-feedback-ui · the "send it?" moment
shadcn-style registry of review-and-edit components for the instant before an agent ships something irreversible. Email composers, SQL runners with READ/WRITE/DESTRUCTIVE badges, patch previews with per-hunk approve/reject, shell command runners with risk badges.
Domain-shaped, not "Approve / Reject."
chessfp · behavioural stylometry from move sequences
Trained a CNN+transformer on 5.7M chess.com decisions to identify which pro you play like. The finding I didn't see coming:
Streamers and pedagogical players are trivial to fingerprint. Elite super-GMs are essentially inseparable from each other. Magnus, Wesley So, Caruana — they converge on engine-approved play and collapse into a single behavioural cluster.
Anna Cramling: 83% accuracy. Magnus Carlsen: 1%. That's a real finding, not a training failure.
claude-md-to-ast-grep · prose conventions, mechanically enforced
Prose conventions in CLAUDE.md rot. ast-grep rules with a teaching note: don't. This skill walks you through converting one into the other, then wires up the post-edit hook so the agent sees its own violations in the next turn.
- mantiq-num2words — multilingual number-to-words for invoices nobody wants to misspell. Handles French 70/80/90, Arabic dual forms.
- blockchain-epitech — ERC-4337 smart account with session keys + EntryPoint v0.7 indexer.
theophilus.homawoo@mantiq.com