this issue: agents that run without a human watching. Sentio i founded and exited, and its Sentinels still sit on deployed contracts and escalate only when something's actually wrong; tracekit mines 8.5M words of my own coding history into a 0.5B model that runs on a laptop; Juriqa orchestrates a bench of legal agents over shared context. page 04 is the one worth reading — it's what i've had to get right to keep any of this running. the trophy shelf is mostly under-25 hackathon hardware, the résumé's on page 07, and the inbox is open (give it a day).
— H.S.
a demo agent is a prompt. a production agent is retries, idempotency, tool schemas, cost ceilings, streaming, human-in-the-loop, and some way to tell whether last week's change made it worse. i build the second kind — multi-agent legal pipelines, autonomous security agents watching live contracts, an LLM reply engine with real users and a billing model, a fine-tuning toolchain that turns agent traces into a local model. some of it i was hired to build; Sentio and tracekit i chose, scoped, and shipped myself. the reason any of it stays up is the boring half of the résumé: Rust services and C++ trading plugins where a dropped message costs money.
built the agentic AI infrastructure from zero. multi-agent orchestration over shared context pipelines with streaming responses, a legal document analysis platform, an LLM reply engine shipped as a Chrome extension with quotas and billing, content pipelines that turn video and PDF into drafts. plus the market-intel side: real-time orderbook tracking and wallet scoring.
elcara.io ↗the whole trading stack, top to bottom. C++ MT5 plugins, Rust services for order and log ingestion, relay servers, analytics, glued together with NATS, Datadog, and onchain invoicing. this is where the reliability instincts come from: everything is async, everything can fail, and someone notices instantly when it does.
whitebeard.ai ↗autonomous security agents for AO smart contracts, and the clearest thing i've owned end to end. i picked the problem, decided the two-layer shape (audit before deploy, Sentinels watching after), wrote the positioning, shipped it, and closed partnerships with Betteridea and Veritas. backed by an Arweave grant and three hackathon placements. then i exited — and the Sentinels are still watching live contracts. i'm comfortable owning the roadmap, not just the ticket.
sentio.hardylabs.cloud ↗a health sensor the size of a coin that hides behind the analog watch you already wear. heart rate, SpO₂, HRV, sleep, no new thing on your wrist. i run the hardware, the brand, and the slightly unreasonable 3D site.
heyvena.com ↗multi-step workflows where each agent gets a narrow tool surface and a typed contract instead of a wall of prompt. shared context pipelines so step four knows what step one found, streaming so the user sees progress instead of a spinner, checkpoints so a failure resumes rather than restarts.
shipped in Juriqa & Viral Pipeline ↓because that's what it is. retries with backoff, idempotent tool calls, schema validation on every structured output, timeouts and token ceilings per run, a dead-letter path for the ones that still fail. same discipline as a trading relay — the model is just a slower, chattier upstream.
learned on Rust relays & NATS ↓a fixed case set, graded runs, and a diff before anything ships. traces captured per run so a bad output can be replayed instead of argued about. vibes are not a regression test, and "it worked when i tried it" is not a deployment signal.
tracekit eval reports ↗frontier models for reasoning, cheap models for routing and extraction, fine-tuned local models where the task is narrow and the volume is high. tracekit takes real agent traces and QLoRA-tunes a 0.5B model that handles ~31% of edit calls offline, at 67 tok/s on a laptop.
github.com/haard18/tracekit ↗Sentio wasn't a ticket handed to me. i chose the problem, shaped it into two layers users could understand, wrote the positioning, shipped it, got the partnerships, and exited it still running. same with tracekit: the hard call was scoping it down — it does imitation SFT on your own traces, not frontier distillation, and the README says so loudly. overclaiming is a product bug.
sentio.hardylabs.cloud ↗Sentio's Sentinels sit on deployed contracts indefinitely, watch for anomalous behaviour, and escalate to a human only when something is actually wrong. weatherbot trades on its own and stands down when expected value doesn't clear the bar. the hard part of autonomy is the not-acting — knowing when to stay quiet and when to wake someone up.
github.com/haard18/weatherbot ↗if you're putting agents in front of real users and need someone who thinks about retries, evals, and cost before the demo, that's my favourite conversation. also happy to talk trading systems, local models, or a hackathon idea that shouldn't work. i read everything. might take a day, but it gets read.